基于改进YOLOv4的道路交通标志识别

基于改进YOLOv4的道路交通标志识别 Road traffic sign recognition algorithm based on improved YOLOv4 first-author 李大湘（1974-），男，湖南人，博士，硕士生导师，主要从事遥感图像分类、目标检测与跟踪、医学图像识别、多实例学习、深度学习等方面的研究。E-mail： www_ldx@163.com http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43440672&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43440703&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43440686&type= 22.01333237 32.00400543 李大湘 （1974-），男，湖南人，博士，硕士生导师，主要从事遥感图像分类、目标检测与跟踪、医学图像识别、多实例学习、深度学习等方面的研究。E-mail： www_ldx@163.com www_ldx@163.com corresp E-mail： Szh1998@stu.xupt.edu.cn E-mail： Szh1998@stu.xupt.edu.cn 苏仲恒（1998-），男，广东肇庆人，硕士研究生，2020年于西安邮电大学获得学士学位，主要从事深度学习、交通场景目标检测与跟踪的研究。E-mail： Szh1998@stu.xupt.edu.cn http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43440704&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43440694&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43440693&type= 22.01333618 32.00399780 苏仲恒 （1998-），男，广东肇庆人，硕士研究生，2020年于西安邮电大学获得学士学位，主要从事深度学习、交通场景目标检测与跟踪的研究。E-mail： Szh1998@stu.xupt.edu.cn Szh1998@stu.xupt.edu.cn 西安邮电大学 通信与信息工程学院，陕西 西安 710121 西安邮电大学 通信与信息工程学院，陕西 西安 710121 College of Communication and Information Engineering， Xi'an University of Posts and Telecommunication， Xi'an 710121， China College of Communication and Information Engineering， Xi'an University of Posts and Telecommunication， Xi'an 710121， China 针对复杂场景中交通标志尺度变化大导致识别精度低的问题，提出了一种改进的YOLOv4算法。首先，设计了一个注意力驱动的尺度感知特征提取模块，通过构建类似残差结构的分层连接方式，增加每层的感受野范围，以获得更具细粒度的多尺度特征，并在注意力驱动下生成一对具有方向感知与位置敏感的注意力图，使网络能聚焦于更具鉴别力的关键区域；然后，构建一个特征对齐的金字塔卷积特征融合模块，即通过卷积计算相邻尺度特征图间的特征偏移量进行特征对齐；最后，通过金字塔卷积的方式使网络自适应学习最优的特征融合模式，并构建特征金字塔用于识别不同尺度的交通标志。实验结果表明，在TT100K数据集上改进算法比原YOLOv4算法的识别精度提高了5.4%，且优于其他对比识别算法，FPS达到33.17，可满足道路交通标志识别的精确性、实时性等要求。 To address the low recognition accuracy resulting from multiple scale changes in the traffic signs of complex scenes， an improved YOLOv4 algorithm is proposed. First， an attention-driven scale-aware feature extraction module is designed， and the range of receptive fields in each layer is widened to obtain more fine-grained multi-scale features by constructing a hierarchical connection mode similar to the residual structure； this is followed by the generation of a pair of attention maps with directional-aware and position-sensitive characteristics under the attention drive so that the network can focus on key areas with more discrimination. Following this， a feature-aligned pyramid convolution feature fusion module is constructed， and the feature offset between adjacent scale feature maps is obtained via convolution for feature alignment. Finally， the network adaptively learns the optimal feature fusion mode through pyramid convolution and constructs a feature pyramid to identify traffic signs with different scales. Experimental results indicate that the recognition accuracy for the TT100K dataset is improved by 5.4% compared with that of the original YOLOv4 algorithm， which is superior to other recognition algorithms， and the FPS reaches 33.17. Thus， the proposed algorithm satisfies the requirements of accuracy and real-time performance for road traffic sign recognition. 计算机视觉 交通标志识别 注意力机制 金字塔卷积 YOLOv4 computer vision traffic sign recognition attention mechanism pyramid convolution YOLOv4 1 引 言 1 车辆作为一种常用的交通工具，在人们的生活中发挥着重要作用。在真实的驾驶场景中，路面交通状况越来越复杂，为了降低交通事故的发生概率，在车辆上安装GPS定位、红外人体感应、雷达测距与摄像头影像采集等设备，再利用无线通信技术构建车联网，是当今智慧交通系统中一种常用的方案 ［ 1 1 - 2 2 1-2 ］ 。交通标志（如限速、停止、禁行等）作为一种通用的交通指挥信号，蕴含着丰富的道路通行控制信息 ［ 3 3 - 4 4 3-4 ］ ，然而，这些信息无法通过传感器直接采集，因此基于车载摄像头获取的影像信息，利用计算机视觉技术对交通标志进行实时识别，是辅助或自动驾驶系统的核心技术，在智能交通中具有重要意义。 根据实现方式不同，交通标志识别算法可以分为两大类：（1）基于手工特征的传统方法，由于交通标志通常在颜色 ［ 5 5 - 6 6 5-6 ］ 与形状 ［ 7 7 - 8 8 7-8 ］ 等方面具有鲜明的物理特征，许多识别算法对此展开研究和设计。由于真实环境下交通标志容易受到光照、遮挡以及运动模糊的影响，手工特征不足以表现不同的交通标志及背景之间的差异，因此存在错误率高的问题，不能满足实际应用需求。（2）基于深度学习的方法。该方法是主流的交通标志识别算法，根据网络结构的不同，可以进一步分为两种：一种是以R-CNN系列为代表的二阶段识别方法 ［ 9 9 - 10 10 9-10 ］ ；另一种是通过去除区域候选阶段的一阶段识别方法，如YOLO ［ 11 11 ］ ，SSD ［ 12 12 ］ 等，它们具有速度快、精度高的优点。 在真实驾驶过程中，车载摄像头与交通标志之间的距离从远至近不断变化，交通标志的尺度也在变化，为了使识别算法具有“多尺度”识别能力，研究人员采用构建图像金字塔或基于特征金字塔网络（Feature Pyramid Network，FPN） ［ 13 13 ］ 的方法构建多尺度网络 ［ 14 14 - 15 15 14-15 ］ 来解决此类问题。Yang等 ［ 16 16 ］ 通过构建图像金字塔，引入Inception模块 ［ 17 17 ］ 设计了DMS-Net尺度感知网络用于交通标志识别。Liu等 ［ 18 18 ］ 采用多尺度反卷积将浅层与深层特征结合形成融合特征图，提出MR-CNN结构，减少了候选区域的数量，从而提升交通标志的检测效率。Liu等 ［ 19 19 ］ 提出的SADA-Net将域自适应网络与多尺度预测网络结合，解决尺度变化的问题。但是，此类方法模型复杂，不具备实时识别能力，在将高维语义信息融入到卷积层时，小尺寸交通标志的特征信息在FPN高层特征融合阶段还会丢失，从而造成对小目标的识别能力严重下降。在车载影像中，占比低于5%的小尺寸交通标志数量众多，而且这些目标缺乏足够的外观信息，要将它们与相似目标或背景区分开来，难度很大。 针对交通标志识别中多尺度与实时性的应用需求，本文对原始YOLOv4 ［ 20 20 ］ 算法进行了改进，设计一个注意力驱动的尺度感知（Attention-driven Scale-aware， ADSA）特征提取残差模块，将它嵌入到CSPDarknet-53特征提取主干网络中，增强网络对多尺度目标的特征提取能力；其次，设计一个特征对齐的金字塔卷积（Feature-aligned Pyramid Convolution， FAPC）特征融合模块，使网络自适应地融合不同尺度的特征，提升网络的特征融合能力，以完成后续交通标志的识别。 2 YOLOv4目标识别算法 1 2.1　YOLOv4网络架构 2 YOLOv4是一个简单高效的一阶段目标检测算法，兼顾实时性与准确性，可将复杂的目标检测问题转换成回归问题进行求解。YOLOv4网络架构如 图1 图1 所示，主要由输入、特征提取骨干网络、颈部特征增强网络以及头部检测网络4部分组成。输入端对输入图片进行等比缩放，将图像尺寸统一（ <math id="M1"><mn mathvariant="normal">608</mn><mo>×</mo><mn mathvariant="normal">608</mn></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441865&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441875&type= 16.00200081 2.79399991 ），并采用不同的数据增强方式，如旋转、基于马赛克（Mosaic）算法的数据增强等，丰富数据集并缩短训练时长，加快算法收敛。 骨干网络在上一代YOLOv3 ［ 21 21 ］ 主干网络Darknet53的基础上，融合CSPNet ［ 22 22 ］ 的网络架构优化思想，组成新的CSPDarknet-53网络作为特征提取网络，这一结构减少了网络的梯度重复信息，增强了骨干网络的特征提取能力，并且显著地减小了网络的参数量。颈部特征增强网络由SPP ［ 23 23 ］ 和PANet ［ 24 24 ］ 组成，形成具备多尺度特征融合能力的特征融合网络结构，极大地减小了模型低级语义信息的流失，进一步提升网络的检测精度。 2.2　YOLOv4识别算法原理 2 YOLOv4识别流程如 图2 图2 所示。YOLOv4算法将图像分成 <math id="M2"><mi>K</mi><mo>×</mo><mi>K</mi></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43440715&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43440719&type= 10.66800022 2.79399991 的网格，目标中心所在的网格负责完成对目标的检测。每个网格需要预测 <math id="M3"><mi>B</mi></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43440722&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43440716&type= 2.53999996 2.79399991 个边界框以及 <math id="M4"><mi>P</mi></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43440738&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43440735&type= 2.62466669 2.79399991 个条件类别概率（ <math id="M5"><mi>P</mi></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43440738&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43440735&type= 2.62466669 2.79399991 为数据集类别数）。每个网格为每个边界框预测中心坐标，预测框的宽高4个值以及存在目标的置信度。最终网络的特征输出维度为 <math id="M6"><mi>K</mi><mo>×</mo><mi>K</mi><mo>×</mo><mn mathvariant="normal">3</mn><mo>×</mo><mo stretchy="false">(</mo><mn mathvariant="normal">5</mn><mo>×</mo><mi>B</mi><mo>+</mo><mi>P</mi><mo stretchy="false">)</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43440742&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43440741&type= 40.89400101 3.55599999 。 YOLOv4算法使用先验框的检测思想，通过K-means聚类方法对数据集中的真实框进行聚类分析，得到大中小3种类别的先验框。边界框的实际位置计算方式如下： <math display="block" id="M7"><mfenced open="{" close="" separators="|"><mrow><mtable columnalign="left"><mtr><mtd><msub><mrow><mi>b</mi></mrow><mrow><mi>x</mi></mrow></msub><mo>=</mo><mi>σ</mi><mo stretchy="false">(</mo><msub><mrow><mi>t</mi></mrow><mrow><mi>x</mi></mrow></msub><mo stretchy="false">)</mo><mo>+</mo><msub><mrow><mi>c</mi></mrow><mrow><mi>x</mi></mrow></msub></mtd></mtr><mtr><mtd><msub><mrow><mi>b</mi></mrow><mrow><mi>y</mi></mrow></msub><mo>=</mo><mi>σ</mi><mo stretchy="false">(</mo><msub><mrow><mi>t</mi></mrow><mrow><mi>y</mi></mrow></msub><mo stretchy="false">)</mo><mo>+</mo><msub><mrow><mi>c</mi></mrow><mrow><mi>y</mi></mrow></msub></mtd></mtr><mtr><mtd><msub><mrow><mi>b</mi></mrow><mrow><mi>w</mi></mrow></msub><mo>=</mo><msub><mrow><mi>p</mi></mrow><mrow><mi>w</mi></mrow></msub><msup><mrow><mi>e</mi></mrow><mrow><msub><mrow><mi>t</mi></mrow><mrow><mi>w</mi></mrow></msub></mrow></msup></mtd></mtr><mtr><mtd><msub><mrow><mi>b</mi></mrow><mrow><mi>h</mi></mrow></msub><mo>=</mo><msub><mrow><mi>p</mi></mrow><mrow><mi>h</mi></mrow></msub><msup><mrow><mi>e</mi></mrow><mrow><msub><mrow><mi>t</mi></mrow><mrow><mi>h</mi></mrow></msub></mrow></msup></mtd></mtr></mtable></mrow></mfenced></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43440766&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43440754&type= 24.72266769 21.50533485 ， 其中： <math id="M8"><mo stretchy="false">(</mo><msub><mrow><mi>t</mi></mrow><mrow><mi>x</mi></mrow></msub><mo>,</mo><msub><mrow><mi>t</mi></mrow><mrow><mi>y</mi></mrow></msub><mo>,</mo><msub><mrow><mi>t</mi></mrow><mrow><mi>w</mi></mrow></msub><mo>,</mo><msub><mrow><mi>t</mi></mrow><mrow><mi>h</mi></mrow></msub><mo stretchy="false">)</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43440759&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43440768&type= 18.54199982 4.48733330 是每个单元格为每个边界框预测出的中心坐标值及框的宽高值， <math id="M9"><mi>σ</mi><mo stretchy="false">(</mo><msub><mrow><mi>t</mi></mrow><mrow><mi>x</mi></mrow></msub><mo stretchy="false">)</mo><mo>,</mo><mi>σ</mi><mo stretchy="false">(</mo><msub><mrow><mi>t</mi></mrow><mrow><mi>y</mi></mrow></msub><mo stretchy="false">)</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43440761&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43440771&type= 17.69533348 4.48733330 是基于矩形框中心点左上角格点坐标的偏移量，通过sigmoid激活函数限制在 <math id="M10"><mo stretchy="false">[</mo><mn mathvariant="normal">0,1</mn><mo stretchy="false">]</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43440779&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43440774&type= 8.97466660 3.55599999 之间， <math id="M11"><msub><mrow><mi>c</mi></mrow><mrow><mi>x</mi></mrow></msub><mo>,</mo><msub><mrow><mi>c</mi></mrow><mrow><mi>y</mi></mrow></msub></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43440781&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43440793&type= 7.28133297 4.14866638 为网格坐标， <math id="M12"><msub><mrow><mi>p</mi></mrow><mrow><mi>w</mi></mrow></msub><mo>,</mo><msub><mrow><mi>p</mi></mrow><mrow><mi>h</mi></mrow></msub></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43440797&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43440796&type= 8.89000034 4.57200003 为先验框的宽高， <math id="M13"><mo stretchy="false">(</mo><msub><mrow><mi>b</mi></mrow><mrow><mi>x</mi></mrow></msub><mo>,</mo><msub><mrow><mi>b</mi></mrow><mrow><mi>y</mi></mrow></msub><mo>,</mo><msub><mrow><mi>b</mi></mrow><mrow><mi>w</mi></mrow></msub><mo>,</mo><msub><mrow><mi>b</mi></mrow><mrow><mi>h</mi></mrow></msub><mo stretchy="false">)</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43440788&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43440786&type= 21.59000015 4.48733330 为预测框实际的坐标以及宽和高。在获得预测框后，通过设置置信度阈值去除低于阈值的预测框，通过DIOU_NMS非极大值抑制后得到预测结果。 3 基于改进YOLOv4的交通标志目标识别 1 3.1　算法整体框架 2 基于YOLOv4算法的一阶段实时交通标志识别网络结构如 图3 图3 所示。首先，以ADSA模块替换原YOLOv4算法中的CSPDarknet53特征提取网络中的残差模块，组成新的注意力驱动的尺度感知特征提取骨干网络ADSA_CSPDarknet。ADSA模块将整个特征作为输入并拆分成组的形式进行处理，通过注意力引导特征学习过程，将此模块嵌入到原始的特征提取骨干网络中使它具备多尺度特征提取能力，且在注意力的引导下使网络聚焦于图像中更重要的区域。其次，通过FAPC模块缓解相邻尺度特征图之间的特征失配问题，自适应地计算相邻尺度间的特征相关性进行特征融合。最后，将融合后的特征送入识别网络完成交通标志识别任务。一阶段方法具有更强大的特征表示能力，并且可以进行端到端的训练，识别速度较快。 3.2　注意力驱动的尺度感知特征提取 2 交通标志在原始图像中的尺度变化范围较大且存在大量小尺寸标志，经过一系列的卷积层后，具有关键作用的特征可能会消失。为了解决这个问题，在Res2Net ［ 25 25 ］ 的启发下，本文设计了一个注意力驱动的多尺度特征提取模块，替换原YOLOv4中的特征提取主干网络CSPDarknet-53中的残差模块。该模块包括多尺度特征提取、通道随机混合以及位置注意力嵌入3个部分。改进后的特征提取主干网络可以提取更具多尺度和细粒度的特征；然后，通过通道随机混合操作，在不增加计算量的条件下更好地融合各通道间的信息；在注意力机制的作用下，在提取特征的过程中保留目标的位置信息，帮助模型更准确定位图像中的感兴趣目标。 3 . 2 . 1　多尺度特征提取 3 注意力驱动的多尺度特征提取模块结构如 图4 图4 所示。与原YOLOv4残差模块相比（见 图1 图1 右下角），通过这种结构可以获得蕴含更丰富的尺度信息的特征表示。区别于原始网络使用 <math id="M14"><mn mathvariant="normal">3</mn><mo>×</mo><mn mathvariant="normal">3</mn></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441398&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441385&type= 8.63599968 2.79399991 卷积核提取特征，ADSA利用一组尺寸较小的 <math id="M15"><mn mathvariant="normal">3</mn><mo>×</mo><mn mathvariant="normal">3</mn></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441398&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441385&type= 8.63599968 2.79399991 卷积核替换单个 <math id="M16"><mn mathvariant="normal">3</mn><mo>×</mo><mn mathvariant="normal">3</mn></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441398&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441385&type= 8.63599968 2.79399991 卷积核，用类似残差结构的方式连接不同组的卷积结果。这样在保持计算量相近的前提下，可以获得多尺度和细粒度更佳的特征表示。具体操作为：在经过 <math id="M17"><mn mathvariant="normal">1</mn><mo>×</mo><mn mathvariant="normal">1</mn></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441472&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441467&type= 8.63599968 2.79399991 卷积层之后，将特征图按通道维度分成 <math id="M18"><mi>n</mi></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43440941&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43440939&type= 1.86266661 2.79399991 组，用 <math id="M19"><msub><mrow><mi>X</mi></mrow><mrow><mi>i</mi></mrow></msub></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43440909&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43440908&type= 3.55599999 3.72533321 表示，其中 <math id="M20"><mi>i</mi><mo>∈</mo><mo stretchy="false">{</mo><mn mathvariant="normal">1,2</mn><mo>,</mo><mo>…</mo><mo>,</mo><mi>n</mi><mo stretchy="false">}</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43440850&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43440848&type= 23.45266724 3.55599999 。与原输入特征图相比，每组特征子集 <math id="M21"><msub><mrow><mi>X</mi></mrow><mrow><mi>i</mi></mrow></msub></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43440909&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43440908&type= 3.55599999 3.72533321 具有相同的空间尺寸，但通道数减少为原来的 <math id="M22"><mn mathvariant="normal">1</mn><mo>/</mo><mi>n</mi></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43440857&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43440869&type= 5.50333309 3.30200005 。每组特征子集都会进行 <math id="M23"><mn mathvariant="normal">3</mn><mo>×</mo><mn mathvariant="normal">3</mn></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43440860&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43440858&type= 8.63599968 2.79399991 卷积操作，用 <math id="M24"><msub><mrow><mi>D</mi></mrow><mrow><mi>i</mi></mrow></msub><mo stretchy="false">(</mo><mo stretchy="false">)</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43440877&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43440863&type= 6.94266701 4.48733330 表示。输出 <math id="M25"><msub><mrow><mi>Y</mi></mrow><mrow><mi>i</mi></mrow></msub></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43440924&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43440923&type= 3.64066648 3.72533321 如下： <math display="block" id="M26"><msub><mrow><mi>Y</mi></mrow><mrow><mi>i</mi></mrow></msub><mo>=</mo><mfenced open="{" close="" separators="|"><mrow><mtable columnalign="left"><mtr><mtd><msub><mrow><mi>D</mi></mrow><mrow><mi>i</mi></mrow></msub><mo stretchy="false">(</mo><msub><mrow><mi>X</mi></mrow><mrow><mi>i</mi></mrow></msub><mo stretchy="false">)</mo><mtext>              </mtext><mi>i</mi><mo>=</mo><mn mathvariant="normal">1</mn></mtd></mtr><mtr><mtd><msub><mrow><mi>D</mi></mrow><mrow><mi>i</mi></mrow></msub><mo stretchy="false">(</mo><msub><mrow><mi>X</mi></mrow><mrow><mi>i</mi></mrow></msub><mo>+</mo><msub><mrow><mi>Y</mi></mrow><mrow><mi>i</mi><mo>-</mo><mn mathvariant="normal">1</mn></mrow></msub><mo stretchy="false">)</mo><mo> </mo><mo> </mo><mo> </mo><mtext>     </mtext><mn mathvariant="normal">1</mn><mo><</mo><mi>i</mi><mo>≤</mo><mi>n</mi></mtd></mtr></mtable></mrow></mfenced></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43440885&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43440882&type= 53.50933075 9.65200043 . 由 式（2） 式（2） 可知，除 <math id="M27"><msub><mrow><mi>X</mi></mrow><mrow><mn mathvariant="normal">1</mn></mrow></msub></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43440893&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43440891&type= 3.89466691 3.72533321 以外，每组特征子集 <math id="M28"><msub><mrow><mi>X</mi></mrow><mrow><mi>i</mi></mrow></msub></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43440909&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43440908&type= 3.55599999 3.72533321 与上一组特征的输出结果相加之后再进行 <math id="M29"><mn mathvariant="normal">3</mn><mo>×</mo><mn mathvariant="normal">3</mn></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441398&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441385&type= 8.63599968 2.79399991 卷积，输出 <math id="M30"><msub><mrow><mi>Y</mi></mrow><mrow><mi>i</mi></mrow></msub></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43440924&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43440923&type= 3.64066648 3.72533321 与 <math id="M31"><msub><mrow><mi>Y</mi></mrow><mrow><mi>j</mi></mrow></msub><mo stretchy="false">(</mo><mi>j</mi><mo><</mo><mi>i</mi><mo stretchy="false">)</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43440918&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43440933&type= 15.07066536 4.82600021 相比具有更大的感受野，因此可以获得不同尺度的子集特征表示。特征图的分组数量 <math id="M32"><mi>n</mi></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43440941&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43440939&type= 1.86266661 2.79399991 越大，可以获得更加丰富的尺度信息，但模型计算量也会随之增加。因此，在衡量模型性能与计算量后， <math id="M33"><mi>n</mi></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43440941&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43440939&type= 1.86266661 2.79399991 的取值为4。将所有输出 <math id="M34"><msub><mrow><mi>Y</mi></mrow><mrow><mi>i</mi></mrow></msub></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43440924&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43440923&type= 3.64066648 3.72533321 沿着通道维度拼接后，采用与ShuffleNet v2 ［ 26 26 ］ 相似的通道随机混合操作，以确保在通道维度上不同组之间的信息流通。最终得到的输出与原输入图像具有相同的尺寸。 通过多尺度特征提取，所学习到的特征包含更深层次的语义特征和浅层的几何特征，增强了全局特征的多样性。因此，本网络可以获得更全面的全局特征表示。 3 . 2 . 2　位置注意力嵌入 3 在交通标志识别任务中，原始图像背景复杂，存在与交通标志相似的虚假目标，并且小尺寸目标的特征信息与位置信息在特征提取阶段极易丢失，因此需要网络具有聚焦于图像中关键区域以及保留目标位置信息的能力。改进的YOLOv4算法引入了新的位置注意力（Coordinate Attention， CA） ［ 27 27 ］ ，CA的结构如 图4 图4 所示。通过在通道注意力中嵌入目标位置信息，在捕获远程依赖关系的同时保留了精确的位置信息，将生成的特征图编码为一对方向敏感和位置敏感的注意力热图，增强目标的表示。具体计算分为两阶段： （1）位置信息嵌入。对于输入 <math id="M35"><mi>X</mi></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43440958&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43440946&type= 2.96333337 2.79399991 ，使用尺寸为 <math id="M36"><mo stretchy="false">(</mo><mi>H</mi><mo>,</mo><mn mathvariant="normal">1</mn><mo stretchy="false">)</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43440961&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43440960&type= 9.73666668 3.55599999 和 <math id="M37"><mo stretchy="false">(</mo><mn mathvariant="normal">1</mn><mo>,</mo><mi>W</mi><mo stretchy="false">)</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43440951&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43440964&type= 10.75266743 3.55599999 的池化核分别沿着水平和垂直方向进行平均池化，对每个通道的位置进行编码。第 <math id="M38"><mi>c</mi></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441680&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441665&type= 1.52400005 2.79399991 个通道中高度为 <math id="M39"><mi>h</mi></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43440976&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43440974&type= 1.94733346 2.79399991 的特征图的输出可以表示为： <math display="block" id="M40"><msubsup><mrow><mi>z</mi></mrow><mrow><mi>c</mi></mrow><mrow><mi>h</mi></mrow></msubsup><mo stretchy="false">(</mo><mi>h</mi><mo stretchy="false">)</mo><mo>=</mo><mfrac><mrow><mn mathvariant="normal">1</mn></mrow><mrow><mi>W</mi></mrow></mfrac><munder><mo movablelimits="false">∑</mo><mrow><mn mathvariant="normal">0</mn><mo>≤</mo><mi>i</mi><mo><</mo><mi>W</mi></mrow></munder><mrow><msub><mrow><mi>x</mi></mrow><mrow><mi>c</mi></mrow></msub><mo stretchy="false">(</mo><mi>h</mi><mo>,</mo><mi>i</mi><mo stretchy="false">)</mo></mrow></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43440981&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43440979&type= 39.45466614 8.89000034 . 同样地，第 <math id="M41"><mi>c</mi></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441680&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441665&type= 1.52400005 2.79399991 个通道中宽度为 <math id="M42"><mi>w</mi></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43440994&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43440992&type= 2.96333337 2.79399991 的特征图的输出可以表示为： <math display="block" id="M43"><msubsup><mrow><mi>z</mi></mrow><mrow><mi>c</mi></mrow><mrow><mi>w</mi></mrow></msubsup><mo stretchy="false">(</mo><mi>w</mi><mo stretchy="false">)</mo><mo>=</mo><mfrac><mrow><mn mathvariant="normal">1</mn></mrow><mrow><mi>H</mi></mrow></mfrac><munder><mo movablelimits="false">∑</mo><mrow><mn mathvariant="normal">0</mn><mo>≤</mo><mi>j</mi><mo><</mo><mi>H</mi></mrow></munder><mrow><msub><mrow><mi>x</mi></mrow><mrow><mi>c</mi></mrow></msub><mo stretchy="false">(</mo><mi>j</mi><mo>,</mo><mi>w</mi><mo stretchy="false">)</mo></mrow></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43440999&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43440996&type= 41.31733322 9.14400005 . 通过上述两个方向上的特征聚合转换，得到一对方向感知的注意力图，使模块同时具有捕获长依赖关系以及保留目标精确位置信息的能力，准确地定位图像中的感兴趣目标。 （2）注意力生成。通过第一阶段的位置信息嵌入后，将两个不同方向的输出 <math id="M44"><msup><mrow><mi>z</mi></mrow><mrow><mi>h</mi></mrow></msup></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441003&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441002&type= 2.96333337 3.04800010 以及 <math id="M45"><msup><mrow><mi>z</mi></mrow><mrow><mi>w</mi></mrow></msup></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441004&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441018&type= 3.47133350 3.04800010 进行拼接，然后送入一个共享参数的 <math id="M46"><mn mathvariant="normal">1</mn><mo>×</mo><mn mathvariant="normal">1</mn></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441472&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441467&type= 8.63599968 2.79399991 卷积转换函数 <math id="M47"><msub><mrow><mi>F</mi></mrow><mrow><mn mathvariant="normal">1</mn></mrow></msub></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441028&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441006&type= 3.72533321 3.72533321 ，数学描述如下： <math display="block" id="M48"><mi>f</mi><mo>=</mo><mi>δ</mi><mo stretchy="false">(</mo><msub><mrow><mi>F</mi></mrow><mrow><mn mathvariant="normal">1</mn></mrow></msub><mo stretchy="false">(</mo><mo stretchy="false">[</mo><msup><mrow><mi>z</mi></mrow><mrow><mi>h</mi></mrow></msup><mo>,</mo><msup><mrow><mi>z</mi></mrow><mrow><mi>w</mi></mrow></msup><mo stretchy="false">]</mo><mo stretchy="false">)</mo><mo stretchy="false">)</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441032&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441030&type= 32.00400162 5.07999992 ， 其中： <math id="M49"><mo stretchy="false">[</mo><mo>⋅</mo><mo>,</mo><mo>⋅</mo><mo stretchy="false">]</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441011&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441010&type= 6.85799980 3.55599999 表示拼接操作， <math id="M50"><mi>δ</mi></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441014&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441012&type= 1.94733346 2.79399991 为非线性激活函数， <math id="M51"><mi>f</mi><mo>∈</mo><msup><mrow><mi mathvariant="bold">R</mi></mrow><mrow><mi>C</mi><mo>/</mo><mi>r</mi><mo>×</mo><mi>H</mi></mrow></msup></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441048&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441016&type= 15.57866764 3.89466691 为沿着水平和垂直方向编码的空间信息的中间特征映射。将 <math id="M52"><mi>f</mi></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441043&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441040&type= 2.45533323 3.64066648 按空间维度分为两个单独的张量： <math id="M53"><msup><mrow><mi mathvariant="bold-italic">f</mi></mrow><mrow><mi>h</mi></mrow></msup><mo>∈</mo><msup><mrow><mi mathvariant="bold">R</mi></mrow><mrow><mi>C</mi><mo>/</mo><mi>r</mi><mo>×</mo><mi>H</mi></mrow></msup></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441054&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441051&type= 17.61066628 3.89466691 和 <math id="M54"><msup><mrow><mi mathvariant="bold-italic">f</mi></mrow><mrow><mi>w</mi></mrow></msup><mo>∈</mo><msup><mrow><mi mathvariant="bold">R</mi></mrow><mrow><mi>C</mi><mo>/</mo><mi>r</mi><mo>×</mo><mi>W</mi></mrow></msup></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441064&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441046&type= 18.54199982 3.89466691 。对两个单独的张量分别使用 <math id="M55"><mn mathvariant="normal">1</mn><mo>×</mo><mn mathvariant="normal">1</mn></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441472&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441467&type= 8.63599968 2.79399991 卷积，将其通道数量转换成原输出数量： <math display="block" id="M56"><msup><mrow><mi>g</mi></mrow><mrow><mi>h</mi></mrow></msup><mo>=</mo><mi>σ</mi><mo stretchy="false">(</mo><msub><mrow><mi>F</mi></mrow><mrow><mi>h</mi></mrow></msub><mo stretchy="false">(</mo><msup><mrow><mi mathvariant="bold-italic">f</mi></mrow><mrow><mi>h</mi></mrow></msup><mo stretchy="false">)</mo><mo stretchy="false">)</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441070&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441061&type= 25.14599991 5.16466665 ， <math display="block" id="M57"><msup><mrow><mi>g</mi></mrow><mrow><mi>w</mi></mrow></msup><mo>=</mo><mi>σ</mi><mo stretchy="false">(</mo><msub><mrow><mi>F</mi></mrow><mrow><mi>w</mi></mrow></msub><mo stretchy="false">(</mo><msup><mrow><mi mathvariant="bold-italic">f</mi></mrow><mrow><mi>w</mi></mrow></msup><mo stretchy="false">)</mo><mo stretchy="false">)</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441074&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441072&type= 26.75466537 5.16466665 ， 其中： <math id="M58"><mi>σ</mi></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441079&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441077&type= 1.86266661 2.79399991 为sigmoid激活函数， <math id="M59"><msub><mrow><mi>F</mi></mrow><mrow><mi>h</mi></mrow></msub></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441098&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441096&type= 3.80999994 3.72533321 和 <math id="M60"><msub><mrow><mi>F</mi></mrow><mrow><mi>w</mi></mrow></msub></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441083&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441100&type= 4.31799984 3.72533321 为两个单独的 <math id="M61"><mn mathvariant="normal">1</mn><mo>×</mo><mn mathvariant="normal">1</mn></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441472&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441467&type= 8.63599968 2.79399991 卷积， <math id="M62"><msup><mrow><mi>g</mi></mrow><mrow><mi>h</mi></mrow></msup></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441090&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441087&type= 3.30200005 3.89466691 与 <math id="M63"><msup><mrow><mi>g</mi></mrow><mrow><mi>w</mi></mrow></msup></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441092&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441106&type= 3.89466691 3.89466691 作为注意力权重。最终CA注意力模块的输出为： <math display="block" id="M64"><msub><mrow><mi>y</mi></mrow><mrow><mi>c</mi></mrow></msub><mo stretchy="false">(</mo><mi>i</mi><mo>,</mo><mi>j</mi><mo stretchy="false">)</mo><mo>=</mo><msub><mrow><mi>x</mi></mrow><mrow><mi>c</mi></mrow></msub><mo stretchy="false">(</mo><mi>i</mi><mo>,</mo><mi>j</mi><mo stretchy="false">)</mo><mo>×</mo><msubsup><mrow><mi>g</mi></mrow><mrow><mi>c</mi></mrow><mrow><mi>h</mi></mrow></msubsup><mo stretchy="false">(</mo><mi>i</mi><mo stretchy="false">)</mo><mo>×</mo><msubsup><mrow><mi>g</mi></mrow><mrow><mi>c</mi></mrow><mrow><mi>w</mi></mrow></msubsup><mo stretchy="false">(</mo><mi>j</mi><mo stretchy="false">)</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441126&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441110&type= 52.57799911 5.16466665 . 将CA注意力嵌入到所设计的多尺度特征提取残差模块中，可以更好地保留图像中的目标标志位置信息，同时使模型聚焦于图像中的真实交通标志区域而忽略相似的虚假标志区域。 3.3　特征对齐金字塔卷积 2 以往的方法在特征融合阶段采用直接按点元素相加或者按通道进行拼接，这种方法往往会忽略相邻特征层之间的相关性，降低模型的性能。由于相邻尺度之间的特征图的相关性更大，所以按照相关度对特征进行融合可以缓解不同特征之间不一致的问题。FAPC模块如 图5 图5 所示，包含自适应金字塔卷积和特征对齐两个部分。 3 . 3 . 1　自适应金字塔卷积 3 以生成特征图 <math id="M65"><msup><mrow><mi>P</mi></mrow><mrow><mn mathvariant="normal">4</mn></mrow></msup></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441212&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441210&type= 3.80999994 3.04800010 为例， <math id="M66"><msup><mrow><mi>F</mi></mrow><mrow><mn mathvariant="normal">3</mn></mrow></msup></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441373&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441372&type= 3.80999994 3.04800010 ， <math id="M67"><msup><mrow><mi>F</mi></mrow><mrow><mn mathvariant="normal">4</mn></mrow></msup></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441139&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441122&type= 3.80999994 3.04800010 ， <math id="M68"><msup><mrow><mi>F</mi></mrow><mrow><mn mathvariant="normal">5</mn></mrow></msup></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441157&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441156&type= 3.80999994 3.04800010 为3个相邻尺度的特征图，利用这3个尺度的特征图融合生成与 <math id="M69"><msup><mrow><mi>F</mi></mrow><mrow><mn mathvariant="normal">4</mn></mrow></msup></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441224&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441222&type= 3.80999994 3.04800010 尺度一致的特征图 <math id="M70"><msup><mrow><mi>P</mi></mrow><mrow><mn mathvariant="normal">4</mn></mrow></msup></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441145&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441164&type= 3.80999994 3.04800010 。对特征对齐后的特征 <math id="M71"><msup><mrow><mi>A</mi></mrow><mrow><mi>n</mi></mrow></msup><mo>,</mo><mi>n</mi><mo>∈</mo><mo stretchy="false">{</mo><mn mathvariant="normal">3,4</mn><mo>,</mo><mn mathvariant="normal">5</mn><mo stretchy="false">}</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441150&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441166&type= 24.29933357 3.80999994 用3个常规二维卷积核构建三维金字塔卷积，最终的输出 <math id="M72"><msup><mrow><mi>P</mi></mrow><mrow><mn mathvariant="normal">4</mn></mrow></msup></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441212&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441210&type= 3.80999994 3.04800010 可以表示为： <math display="block" id="M73"><msup><mrow><mi>P</mi></mrow><mrow><mn mathvariant="normal">4</mn></mrow></msup><mo>=</mo><msub><mrow><mi>W</mi></mrow><mrow><mn mathvariant="normal">0</mn></mrow></msub><mo>⋅</mo><msup><mrow><mi>A</mi></mrow><mrow><mn mathvariant="normal">3</mn></mrow></msup><mo>⋅</mo><msubsup><mrow><mi>λ</mi></mrow><mrow><mn mathvariant="normal">4</mn></mrow><mrow><mn mathvariant="normal">3</mn></mrow></msubsup><mo>+</mo><msub><mrow><mi>W</mi></mrow><mrow><mn mathvariant="normal">1</mn></mrow></msub><mo>⋅</mo><msup><mrow><mi>A</mi></mrow><mrow><mn mathvariant="normal">4</mn></mrow></msup><mo>⋅</mo><msubsup><mrow><mi>λ</mi></mrow><mrow><mn mathvariant="normal">4</mn></mrow><mrow><mn mathvariant="normal">4</mn></mrow></msubsup><mo>+</mo><msub><mrow><mi>W</mi></mrow><mrow><mn mathvariant="normal">2</mn></mrow></msub><mo>⋅</mo><msup><mrow><mi>A</mi></mrow><mrow><mn mathvariant="normal">5</mn></mrow></msup><mo>⋅</mo><msubsup><mrow><mi>λ</mi></mrow><mrow><mn mathvariant="normal">4</mn></mrow><mrow><mn mathvariant="normal">5</mn></mrow></msubsup></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441185&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441184&type= 66.54799652 4.06400013 ， 式中： <math id="M74"><msub><mrow><mi>W</mi></mrow><mrow><mn mathvariant="normal">0</mn></mrow></msub></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441189&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441173&type= 4.99533367 3.72533321 ， <math id="M75"><msub><mrow><mi>W</mi></mrow><mrow><mn mathvariant="normal">1</mn></mrow></msub></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441193&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441191&type= 4.99533367 3.72533321 和 <math id="M76"><msub><mrow><mi>W</mi></mrow><mrow><mn mathvariant="normal">2</mn></mrow></msub></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441196&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441177&type= 4.99533367 3.72533321 是3个二维卷积核，其权重在合成不同尺度特征图时保持一致； <math id="M77"><msubsup><mrow><mi>λ</mi></mrow><mrow><mn mathvariant="normal">4</mn></mrow><mrow><mn mathvariant="normal">3</mn></mrow></msubsup></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441365&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441351&type= 2.96333337 3.97933316 ， <math id="M78"><msubsup><mrow><mi>λ</mi></mrow><mrow><mn mathvariant="normal">4</mn></mrow><mrow><mn mathvariant="normal">4</mn></mrow></msubsup></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441356&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441354&type= 2.96333337 3.97933316 以及 <math id="M79"><msubsup><mrow><mi>λ</mi></mrow><mrow><mn mathvariant="normal">4</mn></mrow><mrow><mn mathvariant="normal">5</mn></mrow></msubsup></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441370&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441357&type= 2.96333337 3.97933316 是相关性系数，代表着 <math id="M80"><msup><mrow><mi>A</mi></mrow><mrow><mi>n</mi></mrow></msup></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441208&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441206&type= 3.97933316 3.04800010 对于 <math id="M81"><msup><mrow><mi>P</mi></mrow><mrow><mn mathvariant="normal">4</mn></mrow></msup></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441212&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441210&type= 3.80999994 3.04800010 尺度特征图的重要性程度， <math id="M82"><mi>n</mi><mo>∈</mo><mo stretchy="false">{</mo><mn mathvariant="normal">3,4</mn><mo>,</mo><mn mathvariant="normal">5</mn><mo stretchy="false">}</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441221&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441220&type= 18.37266731 3.55599999 。 相关系数的计算过程如下：首先，对相邻特征图进行上下采样操作以及特征对齐后获得与 <math id="M83"><msup><mrow><mi>F</mi></mrow><mrow><mn mathvariant="normal">4</mn></mrow></msup></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441224&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441222&type= 3.80999994 3.04800010 分辨率相同的 <math id="M84"><msup><mrow><mi>A</mi></mrow><mrow><mn mathvariant="normal">3</mn></mrow></msup></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441244&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441226&type= 3.97933316 3.04800010 ， <math id="M85"><msup><mrow><mi>A</mi></mrow><mrow><mn mathvariant="normal">4</mn></mrow></msup></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441232&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441245&type= 3.97933316 3.04800010 和 <math id="M86"><msup><mrow><mi>A</mi></mrow><mrow><mn mathvariant="normal">5</mn></mrow></msup></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441236&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441234&type= 3.97933316 3.04800010 ；然后将 <math id="M87"><msup><mrow><mi>A</mi></mrow><mrow><mn mathvariant="normal">3</mn></mrow></msup></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441247&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441246&type= 3.97933316 3.04800010 ， <math id="M88"><msup><mrow><mi>A</mi></mrow><mrow><mn mathvariant="normal">4</mn></mrow></msup></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441253&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441242&type= 3.97933316 3.04800010 和 <math id="M89"><msup><mrow><mi>A</mi></mrow><mrow><mn mathvariant="normal">5</mn></mrow></msup></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441273&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441261&type= 3.97933316 3.04800010 输入到 <math id="M90"><mn mathvariant="normal">1</mn><mo>×</mo><mn mathvariant="normal">1</mn></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441472&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441467&type= 8.63599968 2.79399991 的卷积中，得到3个尺寸为 <math id="M91"><mi>n</mi><mo>×</mo><mi>h</mi><mo>×</mo><mi>w</mi></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441282&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441280&type= 16.42533302 2.79399991 的空间权重向量，并按照通道进行拼接得到 <math id="M92"><mn mathvariant="normal">3</mn><mi>n</mi><mo>×</mo><mi>h</mi><mo>×</mo><mi>w</mi></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441287&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441285&type= 18.20333290 2.79399991 的权重融合图；对权重融合图采用 <math id="M93"><mn mathvariant="normal">1</mn><mo>×</mo><mn mathvariant="normal">1</mn><mo>×</mo><mn mathvariant="normal">3</mn></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441272&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441289&type= 15.32466698 2.79399991 的卷积聚合特征图得到3个尺寸为 <math id="M94"><mi>h</mi><mo>×</mo><mi>w</mi></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441291&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441305&type= 9.65200043 2.79399991 的权重矩阵 <math id="M95"><msubsup><mrow><mi mathvariant="bold-italic">δ</mi></mrow><mrow><mn mathvariant="normal">4</mn></mrow><mrow><mn mathvariant="normal">3</mn></mrow></msubsup></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441294&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441308&type= 3.30200005 3.97933316 ， <math id="M96"><msubsup><mrow><mi mathvariant="bold-italic">δ</mi></mrow><mrow><mn mathvariant="normal">4</mn></mrow><mrow><mn mathvariant="normal">4</mn></mrow></msubsup></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441298&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441311&type= 3.30200005 3.97933316 以及 <math id="M97"><msubsup><mrow><mi mathvariant="bold-italic">δ</mi></mrow><mrow><mn mathvariant="normal">4</mn></mrow><mrow><mn mathvariant="normal">5</mn></mrow></msubsup></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441314&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441300&type= 3.30200005 3.97933316 ；最终，通过softmax函数计算相关性系数，具体计算过程如下： <math display="block" id="M98"><msubsup><mrow><mi>λ</mi></mrow><mrow><mn mathvariant="normal">4</mn></mrow><mrow><mn mathvariant="normal">3</mn></mrow></msubsup><mo stretchy="false">(</mo><mi>i</mi><mo>,</mo><mi>j</mi><mo stretchy="false">)</mo><mo>=</mo><mfrac><mrow><mi mathvariant="normal">e</mi><mi mathvariant="normal">x</mi><mi mathvariant="normal">p</mi><mo stretchy="false">(</mo><msubsup><mrow><mi mathvariant="bold-italic">δ</mi></mrow><mrow><mn mathvariant="normal">4</mn></mrow><mrow><mn mathvariant="normal">3</mn></mrow></msubsup><mo stretchy="false">(</mo><mi>i</mi><mo>,</mo><mi>j</mi><mo stretchy="false">)</mo><mo stretchy="false">)</mo></mrow><mrow><mstyle displaystyle="true"><munderover><mo largeop="true">∑</mo><mrow><mi>n</mi><mo>=</mo><mn mathvariant="normal">3</mn></mrow><mrow><mn mathvariant="normal">5</mn></mrow></munderover></mstyle><mrow><mi mathvariant="normal">e</mi><mi mathvariant="normal">x</mi><mi mathvariant="normal">p</mi><mo stretchy="false">(</mo><msubsup><mrow><mi mathvariant="bold-italic">δ</mi></mrow><mrow><mn mathvariant="normal">4</mn></mrow><mrow><mi>n</mi></mrow></msubsup><mo stretchy="false">(</mo><mi>i</mi><mo>,</mo><mi>j</mi><mo stretchy="false">)</mo><mo stretchy="false">)</mo></mrow></mrow></mfrac></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441320&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441316&type= 41.57133484 13.97000027 ， 式中： <math id="M99"><msubsup><mrow><mi>λ</mi></mrow><mrow><mn mathvariant="normal">4</mn></mrow><mrow><mn mathvariant="normal">3</mn></mrow></msubsup><mo stretchy="false">(</mo><mi>i</mi><mo>,</mo><mi>j</mi><mo stretchy="false">)</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441336&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441334&type= 10.83733368 5.07999992 表示在位置 <math id="M100"><mo stretchy="false">(</mo><mi>i</mi><mo>,</mo><mi>j</mi><mo stretchy="false">)</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441339&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441323&type= 7.11199999 3.97933316 处尺度3与尺度4之间的权重。相关系数 <math id="M101"><mi>λ</mi></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441328&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441341&type= 1.77800000 2.79399991 满足以下关系： <math display="block" id="M102"><msubsup><mrow><mi>λ</mi></mrow><mrow><mn mathvariant="normal">4</mn></mrow><mrow><mn mathvariant="normal">3</mn></mrow></msubsup><mo stretchy="false">(</mo><mi>i</mi><mo>,</mo><mi>j</mi><mo stretchy="false">)</mo><mo>+</mo><msubsup><mrow><mi>λ</mi></mrow><mrow><mn mathvariant="normal">4</mn></mrow><mrow><mn mathvariant="normal">4</mn></mrow></msubsup><mo stretchy="false">(</mo><mi>i</mi><mo>,</mo><mi>j</mi><mo stretchy="false">)</mo><mo>+</mo><msubsup><mrow><mi>λ</mi></mrow><mrow><mn mathvariant="normal">4</mn></mrow><mrow><mn mathvariant="normal">5</mn></mrow></msubsup><mo stretchy="false">(</mo><mi>i</mi><mo>,</mo><mi>j</mi><mo stretchy="false">)</mo><mo>=</mo><mn mathvariant="normal">1</mn></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441345&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441330&type= 47.66733551 5.07999992 ， <math display="block" id="M103"><msubsup><mrow><mi>λ</mi></mrow><mrow><mn mathvariant="normal">4</mn></mrow><mrow><mn mathvariant="normal">3</mn></mrow></msubsup><mo>,</mo><msubsup><mrow><mi>λ</mi></mrow><mrow><mn mathvariant="normal">4</mn></mrow><mrow><mn mathvariant="normal">4</mn></mrow></msubsup><mo>,</mo><msubsup><mrow><mi>λ</mi></mrow><mrow><mn mathvariant="normal">4</mn></mrow><mrow><mn mathvariant="normal">5</mn></mrow></msubsup><mo>∈</mo><mo stretchy="false">[</mo><mn mathvariant="normal">0,1</mn><mo stretchy="false">]</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441347&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441346&type= 25.23066711 4.65666676 . 空间权重 <math id="M104"><msubsup><mrow><mi>λ</mi></mrow><mrow><mn mathvariant="normal">4</mn></mrow><mrow><mn mathvariant="normal">3</mn></mrow></msubsup></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441365&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441351&type= 2.96333337 3.97933316 ， <math id="M105"><msubsup><mrow><mi>λ</mi></mrow><mrow><mn mathvariant="normal">4</mn></mrow><mrow><mn mathvariant="normal">4</mn></mrow></msubsup></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441356&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441354&type= 2.96333337 3.97933316 以及 <math id="M106"><msubsup><mrow><mi>λ</mi></mrow><mrow><mn mathvariant="normal">4</mn></mrow><mrow><mn mathvariant="normal">5</mn></mrow></msubsup></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441370&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441357&type= 2.96333337 3.97933316 可以通过网络的反向传播进行自适应学习，并且在所有通道中共享。由于 <math id="M107"><msup><mrow><mi>F</mi></mrow><mrow><mn mathvariant="normal">3</mn></mrow></msup></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441373&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441372&type= 3.80999994 3.04800010 和 <math id="M108"><msup><mrow><mi>F</mi></mrow><mrow><mn mathvariant="normal">5</mn></mrow></msup></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441378&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441375&type= 3.80999994 3.04800010 为最小尺度和最大尺度，因此只有一层相邻特征层与其对应。最终的模块输出可以改写为： <math display="block" id="M109"><msup><mrow><mi>P</mi></mrow><mrow><mi>l</mi></mrow></msup><mo>=</mo><mfenced open="{" close="" separators="|"><mrow><mtable columnalign="left"><mtr><mtd><msub><mrow><mi>W</mi></mrow><mrow><mn mathvariant="normal">1</mn></mrow></msub><mo>⋅</mo><msup><mrow><mi>A</mi></mrow><mrow><mn mathvariant="normal">3</mn></mrow></msup><mo>⋅</mo><msubsup><mrow><mi>λ</mi></mrow><mrow><mi>l</mi></mrow><mrow><mi>l</mi></mrow></msubsup><mo>+</mo><msub><mrow><mi>W</mi></mrow><mrow><mn mathvariant="normal">2</mn></mrow></msub><mo>⋅</mo><msup><mrow><mi>A</mi></mrow><mrow><mn mathvariant="normal">4</mn></mrow></msup><mo>⋅</mo><msubsup><mrow><mi>λ</mi></mrow><mrow><mi>l</mi></mrow><mrow><mi>l</mi><mo>+</mo><mn mathvariant="normal">1</mn></mrow></msubsup><mo>,</mo><mtext>                      </mtext><mo> </mo><mi>l</mi><mo>=</mo><mn mathvariant="normal">3</mn></mtd></mtr><mtr><mtd><msub><mrow><mi>W</mi></mrow><mrow><mn mathvariant="normal">0</mn></mrow></msub><mo>⋅</mo><msup><mrow><mi>A</mi></mrow><mrow><mn mathvariant="normal">3</mn></mrow></msup><mo>⋅</mo><msubsup><mrow><mi>λ</mi></mrow><mrow><mi>l</mi></mrow><mrow><mi>l</mi><mo>-</mo><mn mathvariant="normal">1</mn></mrow></msubsup><mo>+</mo><msub><mrow><mi>W</mi></mrow><mrow><mn mathvariant="normal">1</mn></mrow></msub><mo>⋅</mo><msup><mrow><mi>A</mi></mrow><mrow><mn mathvariant="normal">4</mn></mrow></msup><mo>⋅</mo><msubsup><mrow><mi>λ</mi></mrow><mrow><mi>l</mi></mrow><mrow><mi>l</mi></mrow></msubsup><mo>+</mo><msub><mrow><mi>W</mi></mrow><mrow><mn mathvariant="normal">2</mn></mrow></msub><mo>⋅</mo><msup><mrow><mi>A</mi></mrow><mrow><mn mathvariant="normal">5</mn></mrow></msup><mo>⋅</mo><msubsup><mrow><mi>λ</mi></mrow><mrow><mi>l</mi></mrow><mrow><mi>l</mi><mo>+</mo><mn mathvariant="normal">1</mn></mrow></msubsup><mo>,</mo><mi>l</mi><mo>=</mo><mn mathvariant="normal">4</mn></mtd></mtr><mtr><mtd><msub><mrow><mi>W</mi></mrow><mrow><mn mathvariant="normal">0</mn></mrow></msub><mo>⋅</mo><msup><mrow><mi>A</mi></mrow><mrow><mn mathvariant="normal">4</mn></mrow></msup><mo>⋅</mo><msubsup><mrow><mi>λ</mi></mrow><mrow><mi>l</mi></mrow><mrow><mi>l</mi><mo>-</mo><mn mathvariant="normal">1</mn></mrow></msubsup><mo>+</mo><msub><mrow><mi>W</mi></mrow><mrow><mn mathvariant="normal">1</mn></mrow></msub><mo>⋅</mo><msup><mrow><mi>A</mi></mrow><mrow><mn mathvariant="normal">5</mn></mrow></msup><mo>⋅</mo><msubsup><mrow><mi>λ</mi></mrow><mrow><mi>l</mi></mrow><mrow><mi>l</mi></mrow></msubsup><mo>,</mo><mtext>                     </mtext><mo> </mo><mo> </mo><mo> </mo><mi>l</mi><mo>=</mo><mn mathvariant="normal">5</mn></mtd></mtr></mtable></mrow></mfenced><mo>.</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441395&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441381&type= 85.76733398 14.90133286 3 . 3 . 2　特征对齐 3 通过上采样操作或步长为2的 <math id="M110"><mn mathvariant="normal">3</mn><mo>×</mo><mn mathvariant="normal">3</mn></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441398&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441385&type= 8.63599968 2.79399991 卷积将特征图 <math id="M111"><msup><mrow><mi>F</mi></mrow><mrow><mi>n</mi><mo>+</mo><mn mathvariant="normal">1</mn></mrow></msup></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441401&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441389&type= 7.36600018 3.04800010 和 <math id="M112"><msup><mrow><mi>F</mi></mrow><mrow><mi>n</mi><mo>-</mo><mn mathvariant="normal">1</mn></mrow></msup></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441404&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441409&type= 7.36600018 3.04800010 的分辨率统一至 <math id="M113"><msup><mrow><mover accent="true"><mi>F</mi><mo>^</mo></mover></mrow><mrow><mi>n</mi></mrow></msup></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441425&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441424&type= 3.80999994 3.97933316 的过程中，特征图之间存在明显的特征空间差异，直接将两层特征图进行融合会导致目标边界模糊，进而影响模型的检测性能。因此，根据 <math id="M114"><msup><mrow><mi>F</mi></mrow><mrow><mi>n</mi></mrow></msup></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441415&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441429&type= 3.80999994 3.04800010 提供的空间位置信息，通过特征对齐模块将相邻的特征图进行特征空间对齐，然后再完成特征融合是很有必要的，特征对齐过程如 图6 图6 所示。 以 <math id="M115"><msup><mrow><mover accent="true"><mi>F</mi><mo>^</mo></mover></mrow><mrow><mi>n</mi></mrow></msup></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441528&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441516&type= 3.80999994 3.97933316 和 <math id="M116"><msup><mrow><mover accent="true"><mi>F</mi><mo>^</mo></mover></mrow><mrow><mi>n</mi><mo>+</mo><mn mathvariant="normal">1</mn></mrow></msup></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441538&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441544&type= 7.36600018 3.97933316 为例，首先将 <math id="M117"><msup><mrow><mover accent="true"><mi>F</mi><mo>^</mo></mover></mrow><mrow><mi>n</mi></mrow></msup><mo>∈</mo><msup><mrow><mi mathvariant="bold">R</mi></mrow><mrow><mi>H</mi><mo>×</mo><mi>W</mi><mo>×</mo><mi>C</mi></mrow></msup></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441451&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441448&type= 20.48933220 4.31799984 和 <math id="M118"><msup><mrow><mover accent="true"><mi>F</mi><mo>^</mo></mover></mrow><mrow><mi>n</mi><mo>+</mo><mn mathvariant="normal">1</mn></mrow></msup><mo>∈</mo><msup><mrow><mi mathvariant="bold">R</mi></mrow><mrow><mi>H</mi><mo>×</mo><mi>W</mi><mo>×</mo><mn mathvariant="normal">2</mn><mi>C</mi></mrow></msup></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441464&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441462&type= 24.97666740 4.31799984 通过 <math id="M119"><mn mathvariant="normal">1</mn><mo>×</mo><mn mathvariant="normal">1</mn></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441472&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441467&type= 8.63599968 2.79399991 卷积将通道数转换为256后进行拼接，然后通过卷积获得两层特征偏移图 <math id="M120"><msub><mrow><mi>Δ</mi></mrow><mrow><msup><mrow><mover accent="true"><mi>F</mi><mo>^</mo></mover></mrow><mrow><mi>n</mi></mrow></msup></mrow></msub><mo>∈</mo><msup><mrow><mi mathvariant="bold">R</mi></mrow><mrow><mi>H</mi><mo>×</mo><mi>W</mi><mo>×</mo><mn mathvariant="normal">2</mn></mrow></msup></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441475&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441473&type= 21.08200073 4.48733330 和 <math id="M121"><msub><mrow><mi>Δ</mi></mrow><mrow><msup><mrow><mover accent="true"><mi>F</mi><mo>^</mo></mover></mrow><mrow><mi>n</mi><mo>+</mo><mn mathvariant="normal">1</mn></mrow></msup></mrow></msub><mo>∈</mo><msup><mrow><mi mathvariant="bold">R</mi></mrow><mrow><mi>H</mi><mo>×</mo><mi>W</mi><mo>×</mo><mn mathvariant="normal">2</mn></mrow></msup></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441480&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441478&type= 23.28333473 4.48733330 。通过特征偏移图 <math id="M122"><msub><mrow><mi>Δ</mi></mrow><mrow><msup><mrow><mi>F</mi></mrow><mrow><mi>n</mi></mrow></msup></mrow></msub></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441500&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441494&type= 4.82600021 3.55599999 和 <math id="M123"><msub><mrow><mi>Δ</mi></mrow><mrow><msup><mrow><mi>F</mi></mrow><mrow><mi>n</mi><mo>-</mo><mn mathvariant="normal">1</mn></mrow></msup></mrow></msub></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441497&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441502&type= 7.02733374 3.55599999 分别将高维特征 <math id="M124"><msup><mrow><mover accent="true"><mi>F</mi><mo>^</mo></mover></mrow><mrow><mi>n</mi><mo>+</mo><mn mathvariant="normal">1</mn></mrow></msup></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441538&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441544&type= 7.36600018 3.97933316 与低维特征 <math id="M125"><msup><mrow><mover accent="true"><mi>F</mi><mo>^</mo></mover></mrow><mrow><mi>n</mi></mrow></msup></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441528&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441516&type= 3.80999994 3.97933316 进行特征对齐。获得特征偏移图后，特征对齐过程如下： <math display="block" id="M126"><msup><mrow><mi>P</mi></mrow><mrow><mi>n</mi></mrow></msup><mo>=</mo><mi mathvariant="bold-italic">U</mi><mfenced separators="|"><mrow><msup><mrow><mover accent="true"><mi>F</mi><mo>^</mo></mover></mrow><mrow><mi>n</mi><mo>+</mo><mn mathvariant="normal">1</mn></mrow></msup><mo>,</mo><msub><mrow><mi>Δ</mi></mrow><mrow><msup><mrow><mover accent="true"><mi>F</mi><mo>^</mo></mover></mrow><mrow><mi>n</mi><mo>+</mo><mn mathvariant="normal">1</mn></mrow></msup></mrow></msub></mrow></mfenced><mo>+</mo><mi mathvariant="bold-italic">U</mi><mfenced separators="|"><mrow><msup><mrow><mover accent="true"><mi>F</mi><mo>^</mo></mover></mrow><mrow><mi>n</mi></mrow></msup><mo>,</mo><msub><mrow><mi>Δ</mi></mrow><mrow><msup><mrow><mover accent="true"><mi>F</mi><mo>^</mo></mover></mrow><mrow><mi>n</mi></mrow></msup></mrow></msub></mrow></mfenced></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441531&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441520&type= 54.18666840 5.92666674 ， 其中 <math id="M127"><mi mathvariant="bold-italic">U</mi><mfenced separators="|"><mrow><mo>⋅</mo><mo>,</mo><mo>⋅</mo></mrow></mfenced></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441526&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441533&type= 11.51466751 3.97933316 为对齐函数。假设特征图 <math id="M128"><msup><mrow><mover accent="true"><mi>F</mi><mo>^</mo></mover></mrow><mrow><mi>n</mi><mo>+</mo><mn mathvariant="normal">1</mn></mrow></msup></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441538&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441544&type= 7.36600018 3.97933316 中需要每个需要对齐的位置空间坐标为 <math id="M129"><mfenced open="{" close="}" separators="|"><mrow><mfenced separators="|"><mrow><mn mathvariant="normal">1,1</mn></mrow></mfenced><mo>,</mo><mfenced separators="|"><mrow><mn mathvariant="normal">1,2</mn></mrow></mfenced><mo>,</mo><mo>.</mo><mo>.</mo><mo>.</mo><mo>,</mo><mfenced separators="|"><mrow><mi>H</mi><mo>,</mo><mi>W</mi></mrow></mfenced></mrow></mfenced></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441550&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441539&type= 45.88933182 4.74133301 ，偏移图 <math id="M130"><mi>Δ</mi><mo>∈</mo><msup><mrow><mi mathvariant="bold">R</mi></mrow><mrow><mi>H</mi><mo>×</mo><mi>W</mi><mo>×</mo><mn mathvariant="normal">2</mn></mrow></msup></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441561&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441551&type= 18.88066673 3.38666677 。 <math id="M131"><msub><mrow><mi>U</mi></mrow><mrow><mi>h</mi><mi>w</mi></mrow></msub></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441564&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441554&type= 5.75733376 3.72533321 为对齐函数 <math id="M132"><mi mathvariant="bold-italic">U</mi><mfenced separators="|"><mrow><mi>P</mi><mo>,</mo><mi>Δ</mi></mrow></mfenced></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441569&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441566&type= 13.71599960 3.97933316 的输出，其定义为： <math display="block" id="M133"><mtable columnalign="left"><mtr><mtd><msub><mrow><mi>U</mi></mrow><mrow><mi>h</mi><mi>w</mi></mrow></msub><mo>=</mo><mstyle displaystyle="true"><munderover><mo largeop="true">∑</mo><mrow><msup><mrow><mi>h</mi></mrow><mrow><mi>'</mi></mrow></msup><mo>=</mo><mn mathvariant="normal">1</mn></mrow><mrow><mi>H</mi></mrow></munderover></mstyle><mrow><mstyle displaystyle="true"><munderover><mo largeop="true">∑</mo><mrow><msup><mrow><mi>w</mi></mrow><mrow><mi>'</mi></mrow></msup><mo>=</mo><mn mathvariant="normal">1</mn></mrow><mrow><mi>W</mi></mrow></munderover></mstyle><mrow><msub><mrow><mover accent="true"><mi>F</mi><mo>^</mo></mover></mrow><mrow><msup><mrow><mi>h</mi></mrow><mrow><mi>'</mi></mrow></msup><msup><mrow><mi>w</mi></mrow><mrow><mi>'</mi></mrow></msup></mrow></msub><mo>⋅</mo><mi mathvariant="normal">m</mi><mi mathvariant="normal">a</mi><mi mathvariant="normal">x</mi><mfenced separators="|"><mrow><mn mathvariant="normal">0,1</mn><mo>-</mo><mfenced open="|" close="|" separators="|"><mrow><mi>h</mi><mo>+</mo><msub><mrow><mi>Δ</mi></mrow><mrow><mn mathvariant="normal">1</mn><mi>h</mi><mi>w</mi></mrow></msub><mo>-</mo><msup><mrow><mi>h</mi></mrow><mrow><mi>'</mi></mrow></msup></mrow></mfenced></mrow></mfenced></mrow></mrow><mo>⋅</mo></mtd></mtr><mtr><mtd><mo> </mo><mo> </mo><mo> </mo><mo> </mo><mo> </mo><mo> </mo><mo> </mo><mo> </mo><mo> </mo><mo> </mo><mo> </mo><mo> </mo><mo> </mo><mo> </mo><mo> </mo><mo> </mo><mo> </mo><mo> </mo><mo> </mo><mo> </mo><mo> </mo><mo> </mo><mo> </mo><mo> </mo><mo> </mo><mi mathvariant="normal">m</mi><mi mathvariant="normal">a</mi><mi mathvariant="normal">x</mi><mfenced separators="|"><mrow><mn mathvariant="normal">0,1</mn><mo>-</mo><mfenced open="|" close="|" separators="|"><mrow><mi>w</mi><mo>+</mo><msub><mrow><mi mathvariant="normal">Δ</mi></mrow><mrow><mn mathvariant="normal">2</mn><mi>h</mi><mi>w</mi></mrow></msub><mo>-</mo><msup><mrow><mi>w</mi></mrow><mrow><mi>'</mi></mrow></msup></mrow></mfenced></mrow></mfenced><mo>,</mo><mtext>               </mtext><mo stretchy="false">(</mo><mn mathvariant="normal">15</mn><mo stretchy="false">)</mo></mtd></mtr></mtable></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441577&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441571&type= 74.59133911 15.15533447 其中：使用双线性插值的方式获得偏移后的位置 <math id="M134"><mfenced separators="|"><mrow><mi>h</mi><mo>+</mo><msub><mrow><mi mathvariant="normal">Δ</mi></mrow><mrow><mn mathvariant="normal">1</mn><mi>h</mi><mi>w</mi></mrow></msub><mo>,</mo><mi>w</mi><mo>+</mo><msub><mrow><mi mathvariant="normal">Δ</mi></mrow><mrow><mn mathvariant="normal">2</mn><mi>h</mi><mi>w</mi></mrow></msub></mrow></mfenced></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441592&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441588&type= 31.91933441 4.06400013 ， <math id="M135"><msub><mrow><mi mathvariant="normal">Δ</mi></mrow><mrow><mn mathvariant="normal">1</mn><mi>h</mi><mi>w</mi></mrow></msub></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441594&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441579&type= 6.18066692 3.64066648 和 <math id="M136"><msub><mrow><mi mathvariant="normal">Δ</mi></mrow><mrow><mn mathvariant="normal">2</mn><mi>h</mi><mi>w</mi></mrow></msub></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441584&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441583&type= 6.18066692 3.64066648 分别表示位置 <math id="M137"><mfenced separators="|"><mrow><mi>h</mi><mo>,</mo><mi>w</mi></mrow></mfenced></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441603&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441585&type= 9.73666668 3.97933316 学习后的二维偏移转换。 相邻尺度的特征图间具有较大相关性的同时也存在明显的语义差别，因此将其直接元素相加或者拼接会导致特征不一致以及目标边界模糊的问题。通过设计的FAPC模块将相邻层间特征图进行特征对齐，通过金字塔卷积的方式使网络自适应学习最优的特征融合模式，更好地完成后续的识别任务。 3.4　损失函数 2 YOLOv4算法的损失函数主要包括3个部分，分别为边界框的置信度损失 <math id="M138"><msub><mrow><mi>L</mi></mrow><mrow><mi mathvariant="normal">C</mi><mi mathvariant="normal">O</mi><mi mathvariant="normal">N</mi><mi mathvariant="normal">F</mi></mrow></msub></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441620&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441618&type= 8.29733276 3.72533321 、分类预测损失 <math id="M139"><msub><mrow><mi>L</mi></mrow><mrow><mi mathvariant="normal">C</mi><mi mathvariant="normal">L</mi><mi mathvariant="normal">A</mi><mi mathvariant="normal">S</mi><mi mathvariant="normal">S</mi></mrow></msub></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441607&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441621&type= 9.14400005 3.72533321 和边界框回归损失 <math id="M140"><msub><mrow><mi>L</mi></mrow><mrow><mi mathvariant="normal">C</mi><mi mathvariant="normal">I</mi><mi mathvariant="normal">o</mi><mi mathvariant="normal">U</mi></mrow></msub></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441610&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441608&type= 7.19666624 3.72533321 。 YOLOv4算法使用CIoU ［ 28 28 ］ 损失作为边界框回归损失函数。CIoU损失函数加入了边界框与预测框之间的中心点距离、重叠面积、宽高比信息，使目标框回归更加稳定。CIoU损失函数的定义如下： <math display="block" id="M141"><msub><mrow><mi>R</mi></mrow><mrow><mi mathvariant="normal">C</mi><mi mathvariant="normal">I</mi><mi mathvariant="normal">o</mi><mi mathvariant="normal">U</mi></mrow></msub><mo>=</mo><mfrac><mrow><msup><mrow><mi>ρ</mi></mrow><mrow><mn mathvariant="normal">2</mn></mrow></msup><mo stretchy="false">(</mo><mi>b</mi><mo>,</mo><msub><mrow><mi>b</mi></mrow><mrow><mi mathvariant="normal">g</mi><mi mathvariant="normal">t</mi></mrow></msub><mo stretchy="false">)</mo></mrow><mrow><msup><mrow><mi>c</mi></mrow><mrow><mn mathvariant="normal">2</mn></mrow></msup></mrow></mfrac><mo>+</mo><mi>α</mi><mi>ν</mi></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441629&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441627&type= 35.22133255 8.46666718 ， <math display="block" id="M142"><mi>ν</mi><mo>=</mo><mfrac><mrow><mn mathvariant="normal">4</mn></mrow><mrow><msup><mrow><mi mathvariant="normal">π</mi></mrow><mrow><mn mathvariant="normal">2</mn></mrow></msup></mrow></mfrac><msup><mrow><mfenced separators="|"><mrow><mi mathvariant="normal">a</mi><mi mathvariant="normal">r</mi><mi mathvariant="normal">c</mi><mi mathvariant="normal">t</mi><mi mathvariant="normal">a</mi><mi mathvariant="normal">n</mi><mfrac><mrow><msup><mrow><mi>w</mi></mrow><mrow><mi mathvariant="normal">g</mi><mi mathvariant="normal">t</mi></mrow></msup></mrow><mrow><msup><mrow><mi>h</mi></mrow><mrow><mi mathvariant="normal">g</mi><mi mathvariant="normal">t</mi></mrow></msup></mrow></mfrac><mo>-</mo><mi mathvariant="normal">a</mi><mi mathvariant="normal">r</mi><mi mathvariant="normal">c</mi><mi mathvariant="normal">t</mi><mi mathvariant="normal">a</mi><mi mathvariant="normal">n</mi><mfrac><mrow><mi>w</mi></mrow><mrow><mi>h</mi></mrow></mfrac></mrow></mfenced></mrow><mrow><mn mathvariant="normal">2</mn></mrow></msup></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441633&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441616&type= 48.93733215 10.32933331 ， <math display="block" id="M143"><msub><mrow><mi>L</mi></mrow><mrow><mi mathvariant="normal">C</mi><mi mathvariant="normal">I</mi><mi mathvariant="normal">o</mi><mi mathvariant="normal">U</mi></mrow></msub><mo>=</mo><mn mathvariant="normal">1</mn><mo>-</mo><mi>I</mi><mi>o</mi><mi>U</mi><mo>+</mo><msub><mrow><mi>R</mi></mrow><mrow><mi mathvariant="normal">C</mi><mi mathvariant="normal">I</mi><mi mathvariant="normal">o</mi><mi mathvariant="normal">U</mi></mrow></msub></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441652&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441651&type= 37.08399963 3.80999994 ， 其中： <math id="M144"><msub><mrow><mi>R</mi></mrow><mrow><mi mathvariant="normal">C</mi><mi mathvariant="normal">I</mi><mi mathvariant="normal">o</mi><mi mathvariant="normal">U</mi></mrow></msub></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441642&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441653&type= 7.53533268 3.72533321 为惩罚项，由最小化预测框与边界框之间的距离定义； <math id="M145"><mi>b</mi></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441657&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441645&type= 1.86266661 2.79399991 和 <math id="M146"><msub><mrow><mi>b</mi></mrow><mrow><mi mathvariant="normal">g</mi><mi mathvariant="normal">t</mi></mrow></msub></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441664&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441647&type= 3.64066648 3.72533321 为预测框与边界框的中心点， <math id="M147"><mi>ρ</mi><mo stretchy="false">(</mo><mo>⋅</mo><mo stretchy="false">)</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441678&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441661&type= 5.16466665 3.89466691 为计算欧氏距离函数， <math id="M148"><mi>c</mi></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441680&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441665&type= 1.52400005 2.79399991 代表能够同时包含预测框和真实框的最小闭区域的对角线距离； <math id="M149"><mi>α</mi></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441710&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441709&type= 1.94733346 2.79399991 为权重系数，使损失函数向重叠区域大的方向优化； <math id="M150"><mi>ν</mi></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441677&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441674&type= 1.43933344 2.79399991 代表预测框与真实框之间宽高比的相似性； <math id="M151"><mo stretchy="false">(</mo><msup><mrow><mi>w</mi></mrow><mrow><mi mathvariant="normal">g</mi><mi mathvariant="normal">t</mi></mrow></msup><mo>,</mo><msup><mrow><mi>h</mi></mrow><mrow><mi mathvariant="normal">g</mi><mi mathvariant="normal">t</mi></mrow></msup><mo stretchy="false">)</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441684&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441695&type= 13.46199989 3.80999994 和 <math id="M152"><mo stretchy="false">(</mo><mi>w</mi><mo>,</mo><mi>h</mi><mo stretchy="false">)</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441691&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441686&type= 9.82133293 3.55599999 分别表示预测边界框和真实边界框的宽度和高度。权重系数 <math id="M153"><mi>α</mi></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441710&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441709&type= 1.94733346 2.79399991 为： <math display="block" id="M154"><mi>α</mi><mo>=</mo><mfrac><mrow><mi>ν</mi></mrow><mrow><mo stretchy="false">(</mo><mn mathvariant="normal">1</mn><mo>-</mo><mi>I</mi><mi>o</mi><mi>U</mi><mo stretchy="false">)</mo><mo>+</mo><mi>ν</mi></mrow></mfrac></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441701&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441698&type= 29.37933159 8.63599968 . 在复杂场景中，交通标志往往只占极小部分的图像区域，因此大部分预设定的锚框只包含复杂的背景区域而不包含交通标志目标，会引起严重的正负样本不均衡问题。改进的YOLOv4算法使用焦点损失函数（Focal Loss，FL） ［ 29 29 ］ 替换原始算法中的交叉熵（Cross-entropy，CE）分类损失函数，通过给不同样本添加权值的方式缓解样本不均衡的问题。CE分类损失函数为： <math display="block" id="M155"><msub><mrow><mi>L</mi></mrow><mrow><mi mathvariant="normal">C</mi><mi mathvariant="normal">E</mi></mrow></msub><mfenced separators="|"><mrow><mi>p</mi><mo>,</mo><mi>y</mi></mrow></mfenced><mo>=</mo><mfenced open="{" close="}" separators="|"><mrow><mtable columnalign="left"><mtr><mtd><mo>-</mo><mi mathvariant="normal">l</mi><mi mathvariant="normal">o</mi><mi mathvariant="normal">g</mi><mo stretchy="false">(</mo><mi>p</mi><mo stretchy="false">)</mo><mo> </mo><mo>,</mo><mo> </mo><mo> </mo><mo> </mo><mo> </mo><mo> </mo><mo> </mo><mo> </mo><mo> </mo><mo> </mo><mo> </mo><mo> </mo><mo> </mo><mo> </mo><mo> </mo><mo> </mo><mo> </mo><mo> </mo><mo> </mo><mo> </mo><mo> </mo><mo> </mo><mo> </mo><mo> </mo><mo> </mo><mo> </mo><mo> </mo><mo> </mo><mo> </mo><mi mathvariant="normal">i</mi><mi mathvariant="normal">f</mi><mo> </mo><mo> </mo><mtext>  </mtext><mi>y</mi><mo>=</mo><mn mathvariant="normal">1</mn></mtd></mtr><mtr><mtd><mo>-</mo><mi mathvariant="normal">l</mi><mi mathvariant="normal">o</mi><mi mathvariant="normal">g</mi><mo stretchy="false">(</mo><mn mathvariant="normal">1</mn><mo>-</mo><mi>p</mi><mo stretchy="false">)</mo><mo>,</mo><mo> </mo><mo> </mo><mo> </mo><mo> </mo><mo> </mo><mo> </mo><mo> </mo><mo> </mo><mo> </mo><mo> </mo><mo> </mo><mo> </mo><mo> </mo><mo> </mo><mo> </mo><mo> </mo><mi mathvariant="normal">o</mi><mi mathvariant="normal">t</mi><mi mathvariant="normal">h</mi><mi mathvariant="normal">e</mi><mi mathvariant="normal">r</mi><mi mathvariant="normal">w</mi><mi mathvariant="normal">i</mi><mi mathvariant="normal">s</mi><mi mathvariant="normal">e</mi></mtd></mtr></mtable></mrow></mfenced></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441717&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441715&type= 65.53199768 10.49866676 ， 其中： <math id="M156"><mi>y</mi><mo>∈</mo><mo stretchy="false">(</mo><mn mathvariant="normal">0,1</mn><mo stretchy="false">)</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441722&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441718&type= 14.22399998 3.97933316 ，为真实标签； <math id="M157"><mi>p</mi><mo>∈</mo><mo stretchy="false">[</mo><mn mathvariant="normal">0,1</mn><mo stretchy="false">]</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441706&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441723&type= 14.98600006 3.97933316 ，表示模型预测类别标签 <math id="M158"><mi>y</mi><mo>=</mo><mn mathvariant="normal">1</mn></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441707&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441727&type= 8.80533314 3.72533321 的概率。FL损失函数是一种改进的分类损失函数，可以表示为： <math display="block" id="M159"><msub><mrow><mi>L</mi></mrow><mrow><mi mathvariant="normal">F</mi><mi mathvariant="normal">L</mi></mrow></msub><mo stretchy="false">(</mo><msub><mrow><mi>p</mi></mrow><mrow><mi>t</mi></mrow></msub><mo stretchy="false">)</mo><mo>=</mo><mo>-</mo><msub><mrow><mi>α</mi></mrow><mrow><mi>t</mi></mrow></msub><mo stretchy="false">(</mo><mn mathvariant="normal">1</mn><mo>-</mo><msub><mrow><mi>p</mi></mrow><mrow><mi>t</mi></mrow></msub><msup><mrow><mo stretchy="false">)</mo></mrow><mrow><mi>γ</mi></mrow></msup><mi mathvariant="normal">l</mi><mi mathvariant="normal">o</mi><mi mathvariant="normal">g</mi><mo stretchy="false">(</mo><msub><mrow><mi>p</mi></mrow><mrow><mi>t</mi></mrow></msub><mo stretchy="false">)</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441731&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441730&type= 49.69933319 5.07999992 ， 其中： <math display="block" id="M160"><msub><mrow><mi>p</mi></mrow><mrow><mi>t</mi></mrow></msub><mo>=</mo><mfenced open="{" close="" separators="|"><mrow><mtable columnalign="left"><mtr><mtd><mi>p</mi><mo> </mo><mo>,</mo><mo> </mo><mo> </mo><mo> </mo><mo> </mo><mo> </mo><mo> </mo><mo> </mo><mo> </mo><mo> </mo><mo> </mo><mo> </mo><mo> </mo><mo> </mo><mo> </mo><mo> </mo><mo> </mo><mo> </mo><mo> </mo><mo> </mo><mo> </mo><mo> </mo><mo> </mo><mo> </mo><mi mathvariant="normal">i</mi><mi mathvariant="normal">f</mi><mo> </mo><mo> </mo><mo> </mo><mo> </mo><mo> </mo><mo> </mo><mi>y</mi><mo>=</mo><mn mathvariant="normal">1</mn></mtd></mtr><mtr><mtd><mn mathvariant="normal">1</mn><mo>-</mo><mi>p</mi><mo>,</mo><mo> </mo><mo> </mo><mo> </mo><mo> </mo><mo> </mo><mo> </mo><mo> </mo><mo> </mo><mo> </mo><mo> </mo><mo> </mo><mi mathvariant="normal">o</mi><mi mathvariant="normal">t</mi><mi mathvariant="normal">h</mi><mi mathvariant="normal">e</mi><mi mathvariant="normal">r</mi><mi mathvariant="normal">w</mi><mi mathvariant="normal">i</mi><mi mathvariant="normal">s</mi><mi mathvariant="normal">e</mi></mtd></mtr></mtable></mrow></mfenced></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441735&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441733&type= 37.76133347 9.90600014 ， 式中： <math id="M161"><msub><mrow><mi>α</mi></mrow><mrow><mi>t</mi></mrow></msub></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441832&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441830&type= 2.70933342 3.72533321 是权重因子，在正负样本之间的取值设定不相同； <math id="M162"><mo stretchy="false">(</mo><mn mathvariant="normal">1</mn><mo>-</mo><msub><mrow><mi>p</mi></mrow><mrow><mi>t</mi></mrow></msub><msup><mrow><mo stretchy="false">)</mo></mrow><mrow><mi>γ</mi></mrow></msup></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441742&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441754&type= 13.46199989 5.07999992 为可变调制系数，通过动态减小易分类样本的权重，使得模型在训练过程中更专注难分类的样本。与CE损失函数相比，FL函数可以提升网络聚焦于负样本的能力。综上，改进YOLOv4的损失函数 <math id="M163"><msub><mrow><mi>L</mi></mrow><mrow><mi mathvariant="normal">l</mi><mi mathvariant="normal">o</mi><mi mathvariant="normal">s</mi><mi mathvariant="normal">s</mi></mrow></msub></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441755&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441743&type= 5.75733376 3.72533321 为： <math display="block" id="M164"><mtable><mtr><mtd><msub><mrow><mi>L</mi></mrow><mrow><mi mathvariant="normal">l</mi><mi mathvariant="normal">o</mi><mi mathvariant="normal">s</mi><mi mathvariant="normal">s</mi></mrow></msub><mo stretchy="false">(</mo><mi>o</mi><mi>b</mi><mi>j</mi><mi>e</mi><mi>c</mi><mi>t</mi><mo stretchy="false">)</mo><mo>=</mo><mstyle displaystyle="true"><munderover><mo largeop="true">∑</mo><mrow><mi>i</mi><mo>=</mo><mn mathvariant="normal">0</mn></mrow><mrow><mi>K</mi><mo>×</mo><mi>K</mi></mrow></munderover></mstyle><mrow><mstyle displaystyle="true"><munderover><mo largeop="true">∑</mo><mrow><mi>j</mi><mo>=</mo><mn mathvariant="normal">0</mn></mrow><mrow><mi>M</mi></mrow></munderover></mstyle><mrow><msubsup><mrow><mi>I</mi></mrow><mrow><mi>i</mi><mi>j</mi></mrow><mrow><mi mathvariant="normal">o</mi><mi mathvariant="normal">b</mi><mi mathvariant="normal">j</mi></mrow></msubsup><mo>⋅</mo></mrow></mrow><msub><mrow><mi>L</mi></mrow><mrow><mi mathvariant="normal">C</mi><mi mathvariant="normal">I</mi><mi mathvariant="normal">o</mi><mi mathvariant="normal">U</mi></mrow></msub><mo>-</mo></mtd></mtr><mtr><mtd><mstyle displaystyle="true"><munderover><mo largeop="true">∑</mo><mrow><mi>i</mi><mo>=</mo><mn mathvariant="normal">0</mn></mrow><mrow><mi>K</mi><mo>×</mo><mi>K</mi></mrow></munderover></mstyle><mrow><mstyle displaystyle="true"><munderover><mo largeop="true">∑</mo><mrow><mi>j</mi><mo>=</mo><mn mathvariant="normal">0</mn></mrow><mrow><mi>M</mi></mrow></munderover></mstyle><mrow><msubsup><mrow><mi>I</mi></mrow><mrow><mi>i</mi><mi>j</mi></mrow><mrow><mi mathvariant="normal">o</mi><mi mathvariant="normal">b</mi><mi mathvariant="normal">j</mi></mrow></msubsup><mo>⋅</mo><msub><mrow><mi>L</mi></mrow><mrow><mi mathvariant="normal">C</mi><mi mathvariant="normal">O</mi><mi mathvariant="normal">N</mi><mi mathvariant="normal">F</mi></mrow></msub></mrow></mrow><mo>+</mo><mo> </mo><msub><mrow><mi>λ</mi></mrow><mrow><mi mathvariant="normal">n</mi><mi mathvariant="normal">o</mi><mi mathvariant="normal">o</mi><mi mathvariant="normal">b</mi><mi mathvariant="normal">j</mi></mrow></msub><mstyle displaystyle="true"><munderover><mo largeop="true">∑</mo><mrow><mi>i</mi><mo>=</mo><mn mathvariant="normal">0</mn></mrow><mrow><mi>K</mi><mo>×</mo><mi>K</mi></mrow></munderover></mstyle><mrow><mstyle displaystyle="true"><munderover><mo largeop="true">∑</mo><mrow><mi>j</mi><mo>=</mo><mn mathvariant="normal">0</mn></mrow><mrow><mi>M</mi></mrow></munderover></mstyle><mrow><msubsup><mrow><mi>I</mi></mrow><mrow><mi>i</mi><mi>j</mi></mrow><mrow><mi mathvariant="normal">n</mi><mi mathvariant="normal">o</mi><mi mathvariant="normal">o</mi><mi mathvariant="normal">b</mi><mi mathvariant="normal">j</mi></mrow></msubsup><mo>⋅</mo><msub><mrow><mi>L</mi></mrow><mrow><mi mathvariant="normal">C</mi><mi mathvariant="normal">O</mi><mi mathvariant="normal">N</mi><mi mathvariant="normal">F</mi></mrow></msub></mrow></mrow><mo>-</mo></mtd></mtr><mtr><mtd><mtext>                            </mtext><mstyle displaystyle="true"><munderover><mo largeop="true">∑</mo><mrow><mi>i</mi><mo>=</mo><mn mathvariant="normal">0</mn></mrow><mrow><mi>K</mi><mo>×</mo><mi>K</mi></mrow></munderover></mstyle><mrow><msubsup><mrow><mi>I</mi></mrow><mrow><mi>i</mi><mi>j</mi></mrow><mrow><mi mathvariant="normal">o</mi><mi mathvariant="normal">b</mi><mi mathvariant="normal">j</mi></mrow></msubsup><munder><mo movablelimits="false">∑</mo><mrow><mi>c</mi><mo>∈</mo><mi mathvariant="normal">c</mi><mi mathvariant="normal">l</mi><mi mathvariant="normal">a</mi><mi mathvariant="normal">s</mi><mi mathvariant="normal">s</mi><mi mathvariant="normal">e</mi><mi mathvariant="normal">s</mi></mrow></munder><mrow><msub><mrow><mi>L</mi></mrow><mrow><mi mathvariant="normal">C</mi><mi mathvariant="normal">L</mi><mi mathvariant="normal">A</mi><mi mathvariant="normal">S</mi><mi mathvariant="normal">S</mi></mrow></msub></mrow></mrow><mo>,</mo><mtext>                      </mtext><mo stretchy="false">(</mo><mn mathvariant="normal">23</mn><mo stretchy="false">)</mo></mtd></mtr></mtable></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441758&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441757&type= 76.53866577 30.31066895 <math display="block" id="M165"><msub><mrow><mi>L</mi></mrow><mrow><mi mathvariant="normal">C</mi><mi mathvariant="normal">O</mi><mi mathvariant="normal">N</mi><mi mathvariant="normal">F</mi></mrow></msub><mo>=</mo><mfenced open="[" close="]" separators="|"><mrow><msub><mrow><mover accent="true"><mi>C</mi><mo>^</mo></mover></mrow><mrow><mi>i</mi></mrow></msub><mi mathvariant="normal">l</mi><mi mathvariant="normal">o</mi><mi mathvariant="normal">g</mi><mo stretchy="false">(</mo><msub><mrow><mi>C</mi></mrow><mrow><mi>i</mi></mrow></msub><mo stretchy="false">)</mo><mo>+</mo><mo stretchy="false">(</mo><mn mathvariant="normal">1</mn><mo>-</mo><msub><mrow><mover accent="true"><mi>C</mi><mo>^</mo></mover></mrow><mrow><mi>i</mi></mrow></msub><mo stretchy="false">)</mo><mi mathvariant="normal">l</mi><mi mathvariant="normal">o</mi><mi mathvariant="normal">g</mi><mo stretchy="false">(</mo><mn mathvariant="normal">1</mn><mo>-</mo><msub><mrow><mi>C</mi></mrow><mrow><mi>i</mi></mrow></msub><mo stretchy="false">)</mo></mrow></mfenced></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441750&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441746&type= 65.95532990 5.92666674 ， <math display="block" id="M166"><mtable columnalign="left"><mtr><mtd><msub><mrow><mi>L</mi></mrow><mrow><mi mathvariant="normal">C</mi><mi mathvariant="normal">L</mi><mi mathvariant="normal">A</mi><mi mathvariant="normal">S</mi><mi mathvariant="normal">S</mi></mrow></msub><mo> </mo><mo>=</mo><mo stretchy="false">(</mo><mn mathvariant="normal">1</mn><mo>-</mo><msub><mrow><mi>α</mi></mrow><mrow><mi>t</mi></mrow></msub><mo stretchy="false">)</mo><msub><mrow><mover accent="true"><mi>p</mi><mo>^</mo></mover></mrow><mrow><mi>i</mi></mrow></msub><msup><mrow><mo stretchy="false">(</mo><mi>c</mi><mo stretchy="false">)</mo></mrow><mrow><mi>γ</mi></mrow></msup><mi mathvariant="normal">l</mi><mi mathvariant="normal">o</mi><mi mathvariant="normal">g</mi><mo stretchy="false">(</mo><mn mathvariant="normal">1</mn><mo>-</mo><msub><mrow><mover accent="true"><mi>p</mi><mo>^</mo></mover></mrow><mrow><mi>i</mi></mrow></msub><mo stretchy="false">(</mo><mi>c</mi><mo stretchy="false">)</mo><mo stretchy="false">)</mo><mo>+</mo></mtd></mtr><mtr><mtd><mtext>                  </mtext><msub><mrow><mi>α</mi></mrow><mrow><mi>t</mi></mrow></msub><mo stretchy="false">(</mo><mn mathvariant="normal">1</mn><mo>-</mo><msub><mrow><mover accent="true"><mi>p</mi><mo>^</mo></mover></mrow><mrow><mi>i</mi></mrow></msub><msup><mrow><mo stretchy="false">(</mo><mi>c</mi><mo stretchy="false">)</mo></mrow><mrow><mi>γ</mi></mrow></msup><mi mathvariant="normal">l</mi><mi mathvariant="normal">o</mi><mi mathvariant="normal">g</mi><mo stretchy="false">(</mo><msub><mrow><mover accent="true"><mi>p</mi><mo>^</mo></mover></mrow><mrow><mi>i</mi></mrow></msub><msup><mrow><mo stretchy="false">(</mo><mi>c</mi><mo stretchy="false">)</mo></mrow><mrow><mi>γ</mi></mrow></msup><mo stretchy="false">)</mo><mo>,</mo><mtext>                  </mtext><mo stretchy="false">(</mo><mn mathvariant="normal">25</mn><mo stretchy="false">)</mo></mtd></mtr></mtable></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441764&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441760&type= 76.62332916 11.17599964 式中：网格序号 <math id="M167"><mi>i</mi><mo>=</mo><mn mathvariant="normal">0,1</mn><mo>,</mo><mo>…</mo><mo>,</mo><mi>K</mi><mo>×</mo><mi>K</mi></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441768&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441766&type= 29.63333130 3.21733332 ；边界框序号 <math id="M168"><mi>j</mi><mo>=</mo><mn mathvariant="normal">0,1</mn><mo>,</mo><mo>…</mo><mo>,</mo><mi>M</mi></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441772&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441770&type= 23.28333473 3.64066648 ； <math id="M169"><msubsup><mrow><mi>I</mi></mrow><mrow><mi>i</mi><mi>j</mi></mrow><mrow><mi mathvariant="normal">o</mi><mi mathvariant="normal">b</mi><mi mathvariant="normal">j</mi></mrow></msubsup></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441774&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441783&type= 4.40266657 3.97933316 表示第 <math id="M170"><mi>i</mi></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441828&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441820&type= 1.10066664 2.79399991 个网格第 <math id="M171"><mi>j</mi></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441795&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441781&type= 1.60866666 3.64066648 个边界框是否负责预测目标，当负责预测目标时取值为1，否则为0； <math id="M172"><msubsup><mrow><mi>I</mi></mrow><mrow><mi>i</mi><mi>j</mi></mrow><mrow><mi mathvariant="normal">n</mi><mi mathvariant="normal">o</mi><mi mathvariant="normal">o</mi><mi mathvariant="normal">b</mi><mi mathvariant="normal">j</mi></mrow></msubsup></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441792&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441790&type= 6.43466663 3.97933316 表示第 <math id="M173"><mi>i</mi></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441828&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441820&type= 1.10066664 2.79399991 个网格第 <math id="M174"><mi>j</mi></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441795&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441781&type= 1.60866666 3.64066648 个边界框是否负责预测目标，当负责预测目标时取值为0，否则为1； <math id="M175"><msub><mrow><mi>λ</mi></mrow><mrow><mi mathvariant="normal">n</mi><mi mathvariant="normal">o</mi><mi mathvariant="normal">o</mi><mi mathvariant="normal">b</mi><mi mathvariant="normal">j</mi></mrow></msub></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441800&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441799&type= 6.34999990 3.72533321 为惩罚权重系数； <math id="M176"><msub><mrow><mi>C</mi></mrow><mrow><mi>i</mi></mrow></msub></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441803&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441802&type= 3.38666677 3.72533321 和 <math id="M177"><msub><mrow><mover accent="true"><mi>C</mi><mo>^</mo></mover></mrow><mrow><mi>i</mi></mrow></msub></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441814&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441804&type= 3.38666677 4.65666676 分别表示第 <math id="M178"><mi>i</mi></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441828&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441820&type= 1.10066664 2.79399991 个网格的置信度预测值和真实值； <math id="M179"><msub><mrow><mi>p</mi></mrow><mrow><mi>i</mi></mrow></msub></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441808&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441817&type= 3.04800010 4.57200003 和 <math id="M180"><msub><mrow><mover accent="true"><mi>p</mi><mo>^</mo></mover></mrow><mrow><mi>i</mi></mrow></msub></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441810&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441819&type= 3.04800010 4.99533367 分别表示第 <math id="M181"><mi>i</mi></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441828&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441820&type= 1.10066664 2.79399991 个网格的类别预测概率值与类别真实值； <math id="M182"><msub><mrow><mi>α</mi></mrow><mrow><mi>t</mi></mrow></msub></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441832&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441830&type= 2.70933342 3.72533321 和 <math id="M183"><mi>γ</mi></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441836&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441835&type= 1.94733346 3.64066648 的取值分别为0.25和2。 4 实验结果及分析 1 4.1　实验数据集 2 为了验证本文所提算法在真实复杂场景下对交通标志的识别性能，使用TT100K［30］交通标志数据集进行实验验证分析。TT100K是一个大型的交通标志数据集，相较于其他交通标志数据集，在信号类别数量以及图像清晰度方面更接近真实复杂驾驶场景。数据集中均为 <math id="M184"><mn mathvariant="normal">2</mn><mtext> </mtext><mn mathvariant="normal">048</mn><mo>×</mo><mn mathvariant="normal">2</mn><mtext> </mtext><mn mathvariant="normal">048</mn></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441841&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441839&type= 21.25133324 2.87866688 的高分辨率图像，包含37 668张训练集图像以及9 417张测试集图像，图像中标注了45类近30 000个交通标志目标，尺寸在 <math id="M185"><mn mathvariant="normal">10</mn><mo>×</mo><mn mathvariant="normal">10</mn></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441822&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441843&type= 12.27666664 2.79399991 ~ <math id="M186"><mn mathvariant="normal">400</mn><mo>×</mo><mn mathvariant="normal">400</mn></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441845&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441823&type= 16.00200081 2.79399991 之间，涵盖不同的交通场景。根据交通标志样本所占的图片像素，可以分为小中大3类，分别为 <math id="M187"><mfenced open="[" close="]" separators="|"><mrow><mfenced separators="|"><mrow><mn mathvariant="normal">0</mn><mo>×</mo><mn mathvariant="normal">0</mn></mrow></mfenced><mo>,</mo><mfenced separators="|"><mrow><mn mathvariant="normal">32</mn><mo>×</mo><mn mathvariant="normal">32</mn></mrow></mfenced></mrow></mfenced></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441847&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441826&type= 32.76599884 4.74133301 、 <math id="M188"><mfenced open="[" close="]" separators="|"><mrow><mfenced separators="|"><mrow><mn mathvariant="normal">32</mn><mo>×</mo><mn mathvariant="normal">32</mn></mrow></mfenced><mo>,</mo><mfenced separators="|"><mrow><mn mathvariant="normal">96</mn><mo>×</mo><mn mathvariant="normal">96</mn></mrow></mfenced></mrow></mfenced></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441850&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441827&type= 36.32199860 4.74133301 以及 <math id="M189"><mfenced open="[" close="]" separators="|"><mrow><mfenced separators="|"><mrow><mn mathvariant="normal">96</mn><mo>×</mo><mn mathvariant="normal">96</mn></mrow></mfenced><mo>,</mo><mfenced separators="|"><mrow><mn mathvariant="normal">400</mn><mo>×</mo><mn mathvariant="normal">400</mn></mrow></mfenced></mrow></mfenced></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441852&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441859&type= 40.13199997 4.74133301 。 图7 图7 展示了TT100K数据集中交通标志的尺寸分布情况。 4.2　实验方法与评价指标 2 实验平台配置如下：GPU为NVIDIA TITANX×4，显存为12 GB，操作系统为Ubuntu 18.04，内存为128 GB，编程语言为Python 3.8，采用Pytorch1.7.0深度学习框架进行算法实现，软件和硬件配置如 表1 表1 所示。 类别 名称 型号和参数 3 硬件 中央处理器 Intel（R）Core（TM） Xeon E5-2640 内存 128GB 图像处理器 NVIDIA Titan X（12G） 4 软件 操作系统 Ubuntu 18.04 深度学习框架 Pytorch-cuda 1.7.0 开发语言 Python 3.8.0 环境管理 Anaconda 3.6.5 模型采用与YOLOv4相同的参数设置策略进行训练调整参数。模型输入图片尺寸为 <math id="M190"><mn mathvariant="normal">608</mn><mo>×</mo><mn mathvariant="normal">608</mn></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441865&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441875&type= 16.00200081 2.79399991 ，初始学习率设定为0.001，batch size为32，整个训练过程为300个epoch，使用余弦退火方法衰减学习率。Anchor设置方面，原YOLOv4算法根据COCO数据集的特点利用K-means聚类算法得到9个不同尺寸的anchor，但COCO数据集目标类别多，尺寸不一，因此其聚类生成的anchor尺寸参数不适用于TT100K数据集。本文利用K-means++算法重新生成anchor用于训练预测，降低原始聚类算法造成的误差。K-means++算法生成的anchor尺寸如 表2 表2 所示。 Feature map Receptive field Anchor box 19×19 Large （74，80） （94，100） （140，146） 38×38 Medium （37，40） （46，52） （58，63） 76×76 Small （14，18） （22，24） （28，32） 为了准确评估模型的检测效果，采用准确率（ P ）和召回率（ R ）指标衡量模型在不同尺度目标下的表现。 <math id="M191"><mi mathvariant="normal">T</mi><mi mathvariant="normal">P</mi></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441868&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441880&type= 5.16466665 2.79399991 （True Positive）表示检测正确的目标， <math id="M192"><mi mathvariant="normal">F</mi><mi mathvariant="normal">P</mi></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441884&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441882&type= 4.74133301 2.79399991 （False Positive）表示检测错误的目标， <math id="M193"><mi mathvariant="normal">F</mi><mi mathvariant="normal">N</mi></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441886&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441870&type= 5.24933338 2.87866688 （False Negative）表示未被检测到的目标，准确率和召回率分别表示为： <math display="block" id="M194"><mi>P</mi><mo>=</mo><mfrac><mrow><mi>T</mi><mi>P</mi></mrow><mrow><mi>T</mi><mi>P</mi><mo>+</mo><mi>F</mi><mi>P</mi></mrow></mfrac></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441890&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441888&type= 22.86000061 8.04333401 ， <math display="block" id="M195"><mi>R</mi><mo>=</mo><mfrac><mrow><mi>T</mi><mi>P</mi></mrow><mrow><mi>T</mi><mi>P</mi><mo>+</mo><mi>F</mi><mi>N</mi></mrow></mfrac></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441871&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441892&type= 23.45266724 8.04333401 . <math id="M196"><mi>F</mi><mn mathvariant="normal">1</mn></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441942&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441940&type= 4.48733330 2.79399991 分数（ <math id="M197"><mi>F</mi><mn mathvariant="normal">1</mn></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441900&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441898&type= 4.48733330 2.79399991 - <math id="M198"><mi mathvariant="normal">s</mi><mi mathvariant="normal">c</mi><mi mathvariant="normal">o</mi><mi mathvariant="normal">r</mi><mi mathvariant="normal">e</mi></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441902&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441901&type= 7.87400007 2.79399991 ）是准确率和召回率的调和平均值。为综合考虑准确率和召回率，使用 <math id="M199"><mi>F</mi><mn mathvariant="normal">1</mn></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441942&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441940&type= 4.48733330 2.79399991 分数对模型进行综合评价，其定义如下： <math display="block" id="M200"><mi>F</mi><mn mathvariant="normal">1</mn><mo>=</mo><mn mathvariant="normal">2</mn><mo>×</mo><mfrac><mrow><mi>P</mi><mo>×</mo><mi>R</mi></mrow><mrow><mi>P</mi><mo>+</mo><mi>R</mi></mrow></mfrac></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441908&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441907&type= 26.67000008 8.04333401 . 4.3　识别结果与分析 2 TT100K拥有更复杂的城市背景，并且交通标志在图像中的占比变化幅度较大，小尺寸标志（如分辨率为 <math id="M201"><mn mathvariant="normal">80</mn><mo>×</mo><mn mathvariant="normal">80</mn></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441923&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441910&type= 12.27666664 2.79399991 ）在原图像中（分辨率为 <math id="M202"><mn mathvariant="normal">2</mn><mtext> </mtext><mn mathvariant="normal">048</mn><mo>×</mo><mn mathvariant="normal">2</mn><mtext> </mtext><mn mathvariant="normal">048</mn></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441914&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441912&type= 21.25133324 2.87866688 ）占比仅有约0.2%，识别难度较大。根据交通标志尺寸，将TT100K数据集分为大中小3个不同类别设计实验，验证改进的YOLOv4算法在交通标志识别任务中的多尺度识别效果，并与YOLOv4 ［ 20 20 ］ ，Zhu等 ［ 30 30 ］ ，Faster R-CNN ［ 31 31 ］ ，FAMN ［ 32 32 ］ ，DR-CNN ［ 33 33 ］ ，Noh等 ［ 34 34 ］ 及Wang等 ［ 35 35 ］ 方法进行比较。 表3 表3 展示了各算法在TT100K数据集上的性能对比结果。首先，改进的YOLOv4算法的平均准确率和 <math id="M203"><mi>F</mi><mn mathvariant="normal">1</mn></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441942&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441940&type= 4.48733330 2.79399991 分数分别达到了93.7%和94.1%，相较于原YOLOv4算法的平均准确率与 <math id="M204"><mi>F</mi><mn mathvariant="normal">1</mn></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441942&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441940&type= 4.48733330 2.79399991 分数分别提升了5.4%和4.8%，在大、中、小目标的准确率分别达到了96.4%，95.5%，89.4%，准确率与原始YOLOv4算法相比分别提升了5.8%，4.2%，6.4%，说明改进后的算法更适合交通标志识别任务，并且在小尺寸目标的提升较为明显。其次，改进算法的平均 <math id="M205"><mi>F</mi><mn mathvariant="normal">1</mn></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441942&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441940&type= 4.48733330 2.79399991 分数为94.1%，在大、中、小目标的 <math id="M206"><mi>F</mi><mn mathvariant="normal">1</mn></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441942&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441940&type= 4.48733330 2.79399991 分数分别达到了96.9%，96.1%，90.8%，说明算法在准确率与召回率之间基本取得了平衡。与TsingNet相比，小目标准确率提升了0.4%，并且中型以及大型目标有小幅提升，说明改进算法在提升小型目标识别性能的同时具备多尺度识别能力。通过多尺度特征提取方式，改进方法在小目标准确率提升上取得最佳效果（89.4%）；与其他识别方法相比，在位置注意力的作用下，可以保留更加精准的目标位置信息，对于在复杂交通环境下的交通标志识别是至关重要的。通过自动学习最优特征融合方式，将高维语义信息与低维空间信息融合，增强多尺度提取能力，与其他方法相比总体的识别准确率最高（93.7%）。为了更直观展示改进YOLOv4算法的效果， 图8 图8 展示了4组与原YOLOv4算法对比的测试结果。可以看到，改进YOLOv4算法可以正确识别出不同尺度的目标，而原YOLOv4算法对于小尺寸交通标志目标则存在漏检的情况（前2组）。改进YOLOv4算法在FAPC模块的作用下，通过特征对齐可以获得更精确的目标边缘信息，边界框回归更准确，而原YOLOv4算法虽然可以识别目标，但是其边界框存在较大偏差，目标定位不准确（第3组）。当图像中背景更为复杂，并且存在遮挡的情况下，改进YOLOv4算法仍可正确识别图像中的目标（第4组），由此表明，本文所提算法在真实驾驶场景中性能更佳。 2 Methods 3 Small 3 Medium 3 Large 3 Overall P R F P R F P R F P R F Zhu et al . ［ 32 32 ］ 81.7 87.4 84.5 90.8 93.6 92.2 90.6 87.7 89.1 87.7 89.6 88.6 Faster R-CNN ［ 33 33 ］ 25.2 58.2 35.2 63.8 83.7 72.4 80.7 91.2 85.6 56.7 77.6 65.5 YOLOv4 ［ 20 20 ］ 83.0 87.5 85.2 91.3 95.2 93.2 90.6 88.2 89.4 88.3 90.3 89.3 FAMN ［ 34 34 ］ 88.4 90.1 89.2 94.2 97.2 95.7 92.8 96.1 94.4 91.8 94.3 93.0 DR-CNN ［ 35 35 ］ 83.1 89.3 86.1 91.7 94.8 93.2 92.4 89.6 91.0 89.0 91.2 90.1 Noh et al . ［36］ 84.8 92.6 88.5 94.2 97.5 95.8 93.3 97.5 95.4 90.6 95.7 93.1 Wang et al . ［37］ 87.3 89.4 88.3 92.5 96.4 94.4 92.8 90.5 91.6 90.8 92.3 91.5 TsingNet ［ 15 15 ］ 89.0 90.6 89.8 95.2 95.6 95.4 96.2 92.8 94.5 93.4 92.9 93.1 Ours 89.4 92.3 90.8 95.5 96.8 96.1 96.4 97.4 96.9 93.7 94.5 94.1 % 4.4　消融实验 2 为了验证改进YOLOv4算法关键模块在交通标志识别任务中的有效性，设置了一组消融实验，分别使用ADSA模块和FAPC模块与原始YOLOv4方法进行对比测试分析。原YOLOv4使用CSPDarkNet-53作为特征提取网络，采用PANet进行特征融合。在整个实验过程中，训练设置包括数据增强方式，模型初始化，学习率与迭代次数等超参数均保持一致。消融实验结果如 表4 表4 所示，其中， <math id="M207"><msub><mrow><mi>P</mi></mrow><mrow><mi mathvariant="normal">s</mi></mrow></msub></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441930&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441945&type= 3.55599999 3.72533321 ， <math id="M208"><msub><mrow><mi>P</mi></mrow><mrow><mi mathvariant="normal">m</mi></mrow></msub></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441951&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441948&type= 4.31799984 3.72533321 ， <math id="M209"><msub><mrow><mi>P</mi></mrow><mrow><mi mathvariant="normal">l</mi></mrow></msub></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441963&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441932&type= 3.21733332 3.72533321 以及 <math id="M210"><msub><mrow><mi>P</mi></mrow><mrow><mi mathvariant="normal">a</mi><mi mathvariant="normal">l</mi><mi mathvariant="normal">l</mi></mrow></msub></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441966&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441965&type= 4.65666676 3.72533321 分别表示小、中、大目标以及整体的准确率。 Methods ADSA FAPC <math id="M211"><msub><mrow><mi>P</mi></mrow><mrow><mi mathvariant="normal">s</mi></mrow></msub></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441970&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441968&type= 3.04800010 3.21733332 <math id="M212"><msub><mrow><mi>P</mi></mrow><mrow><mi mathvariant="normal">m</mi></mrow></msub></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441954&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441972&type= 3.72533321 3.21733332 <math id="M213"><msub><mrow><mi>P</mi></mrow><mrow><mi mathvariant="normal">l</mi></mrow></msub></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441959&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441957&type= 2.79399991 3.21733332 <math id="M214"><msub><mrow><mi>P</mi></mrow><mrow><mi mathvariant="normal">a</mi><mi mathvariant="normal">l</mi><mi mathvariant="normal">l</mi></mrow></msub></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441962&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441960&type= 3.97933316 3.21733332 FPS YOLOv4 baseline - - 83.0 91.3 90.6 88.3 41.36 <math id="M215"><mo>√</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441997&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441982&type= 2.70933342 2.96333337 - 87.3 92.7 93.4 91.2 37.74 - <math id="M216"><mo>√</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441997&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441982&type= 2.70933342 2.96333337 86.6 93.8 94.5 91.6 38.32 Ours <math id="M217"><mo>√</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441997&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441982&type= 2.70933342 2.96333337 <math id="M218"><mo>√</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441997&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=43441982&type= 2.70933342 2.96333337 89.4 95.5 96.4 93.7 33.17 % 从 表4 表4 可以看出：所设计模块对于原始算法精度均有提升，将两个模块组合使用得到的提升效果最大，原YOLOv4的平均准确率为88.3%，组合使用两个改进模块后的YOLOv4模型准确率达到了93.7%，提升了5.4%；应用ADSA模块后小目标准确率提升较大，达到了87.3%，说明ADSA模块通过类似残差连接的多尺度特征提取方式以及位置注意力的作用下，扩大了感受野范围，可以很好地关注到图像的重要区域，提取到更精确的特征并且更好地保留小目标的位置信息，以完成交通标志识别；FAPC模块对于各尺度目标的识别准确率均有提升，通过构建特征金字塔，将高层语义信息与低维空间信息结合，增强识别效果；相较于原YOLOv4算法，应用两个设计模块后的模型识别速度为33.17，更适合在真实场景下完成交通标志识别任务。 5 结 论 1 针对真实场景下交通标志识别任务背景复杂，目标尺度变化大的挑战，本文改进了YOLOv4算法，通过使用多尺度特征提取方式以及嵌入注意力机制，改进后的网络可以更好地关注图像中的重要区域并增强小目标交通标志的特征表示能力。其次，设计了一个新的特征融合模块，使网络充分利用不同尺度的特征信息完成多尺度识别。改进算法在TT100K数据集上的平均准确率和召回率分别为93.7%和94.5%，FPS达到33.17，能够实现真实场景下交通标志的准确识别。后续研究将从进一步提升模型识别速度出发，改进优化模型网络结构，以提升识别效率。 参考文献 KUUTTI S ， FALLAH S ， KATSAROS K ， et al . A survey of the state-of-the-art localization techniques and their potentials for autonomous vehicle applications ［J］. IEEE Internet of Things Journal ， 2018 ， 5 （ 2 ）： 829 - 846 . doi: 10.1109/jiot.2018.2812300 http://dx.doi.org/10.1109/jiot.2018.2812300 A survey of the state-of-the-art localization techniques and their potentials for autonomous vehicle applications ZHENG Y ， WANG J W ， LI K Q . Smoothing traffic flow via control of autonomous vehicles ［J］. IEEE Internet of Things Journal ， 2020 ， 7 （ 5 ）： 3882 - 3896 . doi: 10.1109/jiot.2020.2966506 http://dx.doi.org/10.1109/jiot.2020.2966506 Smoothing traffic flow via control of autonomous vehicles YANG X H ， LIU W F ， ZHANG S L ， et al . Targeted attention attack on deep learning models in road sign recognition ［J］. IEEE Internet of Things Journal ， 2021 ， 8 （ 6 ）： 4980 - 4990 . doi: 10.1109/jiot.2020.3034899 http://dx.doi.org/10.1109/jiot.2020.3034899 Targeted attention attack on deep learning models in road sign recognition LI Y J ， XU X ， XIAO J H ， et al . Adaptive Square attack： fooling autonomous cars with adversarial traffic signs ［J］. IEEE Internet of Things Journal ， 2021 ， 8 （ 8 ）： 6337 - 6347 . doi: 10.1109/jiot.2020.3016145 http://dx.doi.org/10.1109/jiot.2020.3016145 Adaptive Square attack： fooling autonomous cars with adversarial traffic signs AKATSUKA H ， IMAI S . Road signposts recognition system ［C］. SAE Technical Paper Series. 400 Commonwealth Drive， Warrendale， PA， United States ： SAE International ， 1987 . doi: 10.4271/870239 http://dx.doi.org/10.4271/870239 Road signposts recognition system MALDONADO-BASCON S ， LAFUENTE-ARROYO S ， GIL-JIMENEZ P ， et al . Road-sign detection and recognition based on support vector machines ［J］. IEEE Transactions on Intelligent Transportation Systems ， 2007 ， 8 （ 2 ）： 264 - 278 . doi: 10.1109/tits.2007.895311 http://dx.doi.org/10.1109/tits.2007.895311 Road-sign detection and recognition based on support vector machines Kuo W J ， Lin C C . Two-stage road sign detection and recognition ［C］. 2007 IEEE International Conference on Multimedia and Expo . IEEE ， 2007 ： 1427 - 1430 . doi: 10.1109/icme.2007.4284928 http://dx.doi.org/10.1109/icme.2007.4284928 Two-stage road sign detection and recognition MADANI A ， YUSOF R . Traffic sign recognition based on color， shape， and pictogram classification using support vector machines ［J］. Neural Computing and Applications ， 2018 ， 30 ： 2807 - 2817 . doi: 10.1007/s00521-017-2887-x http://dx.doi.org/10.1007/s00521-017-2887-x Traffic sign recognition based on color， shape， and pictogram classification using support vector machines LI J ， WANG Z F . Real-time traffic sign recognition based on efficient CNNs in the wild ［J］. IEEE Transactions on Intelligent Transportation Systems ， 2019 ， 20 （ 3 ）： 975 - 984 . doi: 10.1109/tits.2018.2843815 http://dx.doi.org/10.1109/tits.2018.2843815 Real-time traffic sign recognition based on efficient CNNs in the wild ZHANG J M ， XIE Z P ， SUN J ， et al . A cascaded R-CNN with multiscale attention and imbalanced samples for traffic sign detection ［J］. IEEE Access ， 2020 ， 8 ： 29742 - 29754 . doi: 10.1109/access.2020.2972338 http://dx.doi.org/10.1109/access.2020.2972338 A cascaded R-CNN with multiscale attention and imbalanced samples for traffic sign detection ZHANG J M ， HUANG M T ， JIN X K ， et al . A real-time Chinese traffic sign detection algorithm based on modified YOLOv2 ［J］. Algorithms ， 2017 ， 10 （ 4 ）： 127 . doi: 10.3390/a10040127 http://dx.doi.org/10.3390/a10040127 A real-time Chinese traffic sign detection algorithm based on modified YOLOv2 SHAN H X ， ZHU W X . A small traffic sign detection algorithm based on modified SSD ［J］. IOP Conference Series： Materials Science and Engineering ， 2019 ， 646 （ 1 ）： 012006 . doi: 10.1088/1757-899x/646/1/012006 http://dx.doi.org/10.1088/1757-899x/646/1/012006 A small traffic sign detection algorithm based on modified SSD LIN T Y ， DOLLÁR P ， GIRSHICK R ， et al . Feature pyramid networks for object detection ［C］. 2017 IEEE Conference on Computer Vision and Pattern Recognition （CVPR）. 2126，2017 ， Honolulu， HI， USA. IEEE ， 2017 ： 936 - 944 . doi: 10.1109/cvpr.2017.106 http://dx.doi.org/10.1109/cvpr.2017.106 Feature pyramid networks for object detection YUAN Y ， XIONG Z T ， WANG Q . VSSA-NET： vertical spatial sequence attention network for traffic sign detection ［J］. IEEE Transactions on Image Processing： a Publication of the IEEE Signal Processing Society ， 2019 ， 28 （ 7 ）： 3423 - 3434 . doi: 10.1109/tip.2019.2896952 http://dx.doi.org/10.1109/tip.2019.2896952 VSSA-NET： vertical spatial sequence attention network for traffic sign detection LIU Y Y ， PENG J Y ， XUE J H ， et al . TSingNet： Scale-aware and context-rich feature learning for traffic sign detection and recognition in the wild ［J］. Neurocomputing ， 2021 ， 447 ： 10 - 22 . doi: 10.1016/j.neucom.2021.03.049 http://dx.doi.org/10.1016/j.neucom.2021.03.049 TSingNet： Scale-aware and context-rich feature learning for traffic sign detection and recognition in the wild YANG Y ， LIU S ， MA W ， et al . Efficient traffic-sign recognition with scale-aware CNN ［EB/OL］. 2018 ： arXiv ： 1805 . 12289 . https：//arxiv.org/abs/1805.12289 https://arxiv.org/abs/1805.12289 . doi: 10.5244/c.31.168 http://dx.doi.org/10.5244/c.31.168 Efficient traffic-sign recognition with scale-aware CNN SZEGEDY C ， LIU W ， JIA Y Q ， et al . Going deeper with convolutions ［C］. 2015 IEEE Conference on Computer Vision and Pattern Recognition （CVPR）. 712，2015 ， Boston， MA， USA. IEEE ， 2015 ： 1 - 9 . doi: 10.1109/cvpr.2015.7298594 http://dx.doi.org/10.1109/cvpr.2015.7298594 Going deeper with convolutions LIU Z G ， DU J ， TIAN F ， et al . MR-CNN： a multi-scale region-based convolutional neural network for small traffic sign recognition ［J］. IEEE Access ， 2019 ， 7 ： 57120 - 57128 . doi: 10.1109/access.2019.2913882 http://dx.doi.org/10.1109/access.2019.2913882 MR-CNN： a multi-scale region-based convolutional neural network for small traffic sign recognition LIU Z W ， SHEN C ， QI M Y ， et al . SADANet： integrating scale-aware and domain adaptive for traffic sign detection ［J］. IEEE Access ， 2020 ， 8 ： 77920 - 77933 . doi: 10.1109/access.2020.2989758 http://dx.doi.org/10.1109/access.2020.2989758 SADANet： integrating scale-aware and domain adaptive for traffic sign detection BOCHKOVSKIY A ， WANG C Y ， LIAO H Y M . YOLOv4： optimal speed and accuracy of object detection ［EB/OL］. 2020 ： arXiv ： 2004 . 10934 . https：//arxiv.org/abs/2004.10934 https://arxiv.org/abs/2004.10934 . YOLOv4： optimal speed and accuracy of object detection REDMON J ， FARHADI A . YOLOv3： an incremental improvement ［EB/OL］. 2018 ： arXiv ： 1804 . 02767 . https：//arxiv.org/abs/1804.02767 https://arxiv.org/abs/1804.02767 . doi: 10.1109/cvpr.2017.690 http://dx.doi.org/10.1109/cvpr.2017.690 YOLOv3： an incremental improvement WANG C Y ， MARK LIAO H Y ， WU Y H ， et al . CSPNet： a new backbone that can enhance learning capability of CNN ［C］. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops （CVPRW）. 1419，2020 ， Seattle， WA， USA. IEEE ， 2020 ： 1571 - 1580 . doi: 10.1109/cvprw50498.2020.00203 http://dx.doi.org/10.1109/cvprw50498.2020.00203 CSPNet： a new backbone that can enhance learning capability of CNN HE K M ， ZHANG X Y ， REN S Q ， et al . Spatial pyramid pooling in deep convolutional networks for visual recognition ［J］. IEEE Transactions on Pattern Analysis and Machine Intelligence ， 2015 ， 37 （ 9 ）： 1904 - 1916 . doi: 10.1109/tpami.2015.2389824 http://dx.doi.org/10.1109/tpami.2015.2389824 Spatial pyramid pooling in deep convolutional networks for visual recognition LIU S ， QI L ， QIN H F ， et al . Path aggregation network for instance segmentation ［C］. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition . 1823，2018 ， Salt Lake City， UT， USA . IEEE ， 2018 ： 8759 - 8768 . doi: 10.1109/cvpr.2018.00913 http://dx.doi.org/10.1109/cvpr.2018.00913 Path aggregation network for instance segmentation GAO S H ， CHENG M M ， ZHAO K ， et al . Res2Net： a new multi-scale backbone architecture ［J］. IEEE Transactions on Pattern Analysis and Machine Intelligence ， 2021 ， 43 （ 2 ）： 652 - 662 . doi: 10.1109/tpami.2019.2938758 http://dx.doi.org/10.1109/tpami.2019.2938758 Res2Net： a new multi-scale backbone architecture MA N ， ZHANG X ， ZHENG H T ， et al . Shufflenet v2： Practical guidelines for efficient cnn architecture design ［C］. Proceedings of the European conference on computer vision （ECCV） . 2018 ： 116 - 131 . doi: 10.1007/978-3-030-01264-9_8 http://dx.doi.org/10.1007/978-3-030-01264-9_8 Shufflenet v2： Practical guidelines for efficient cnn architecture design HOU Q B ， ZHOU D Q ， FENG J S . Coordinate attention for efficient mobile network design ［C］. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition （CVPR）. 2025，2021 ， Nashville， TN， USA. IEEE ， 2021 ： 13708 - 13717 . doi: 10.1109/cvpr46437.2021.01350 http://dx.doi.org/10.1109/cvpr46437.2021.01350 Coordinate attention for efficient mobile network design ZHENG Z H ， WANG P ， LIU W ， et al . Distance-IoU loss： faster and better learning for bounding box regression ［J］. Proceedings of the AAAI Conference on Artificial Intelligence ， 2020 ， 34 （ 7 ）： 12993 - 13000 . doi: 10.1609/aaai.v34i07.6999 http://dx.doi.org/10.1609/aaai.v34i07.6999 Distance-IoU loss： faster and better learning for bounding box regression LIN T Y ， GOYAL P ， GIRSHICK R ， et al . Focal loss for dense object detection ［C］. 2017 IEEE International Conference on Computer Vision （ICCV）. 2229，2017 ， Venice， Italy. IEEE ， 2017 ： 2999 - 3007 . doi: 10.1109/iccv.2017.324 http://dx.doi.org/10.1109/iccv.2017.324 Focal loss for dense object detection ZHU Z ， LIANG D ， ZHANG S H ， et al . Traffic-sign detection and classification in the wild ［C］. 2016 IEEE Conference on Computer Vision and Pattern Recognition （CVPR）. 2730，2016 ， Las Vegas， NV， USA. IEEE ， 2016 ： 2110 - 2118 . doi: 10.1109/cvpr.2016.232 http://dx.doi.org/10.1109/cvpr.2016.232 Traffic-sign detection and classification in the wild REN S Q ， HE K M ， GIRSHICK R ， et al . Faster R-CNN： towards real-time object detection with region proposal networks ［J］. IEEE Transactions on Pattern Analysis and Machine Intelligence ， 2017 ， 39 （ 6 ）： 1137 - 1149 . doi: 10.1109/tpami.2016.2577031 http://dx.doi.org/10.1109/tpami.2016.2577031 Faster R-CNN： towards real-time object detection with region proposal networks OU Z H ， XIAO F R ， XIONG B Q ， et al . FAMN： feature aggregation multipath network for small traffic sign detection ［J］. IEEE Access ， 2019 ， 7 ： 178798 - 178810 . doi: 10.1109/access.2019.2959015 http://dx.doi.org/10.1109/access.2019.2959015 FAMN： feature aggregation multipath network for small traffic sign detection LIU Z G ， LI D Y ， GE S S ， et al . Small traffic sign detection from large image ［J］. Applied Intelligence ， 2020 ， 50 （ 1 ）： 1 - 13 . doi: 10.1007/s10489-019-01511-7 http://dx.doi.org/10.1007/s10489-019-01511-7 Small traffic sign detection from large image NOH J ， BAE W ， LEE W ， et al . Better to follow， follow to be better： towards precise supervision of feature super-resolution for small object detection ［C］. 2019 IEEE/CVF International Conference on Computer Vision （ICCV）. 272，2019 ， Seoul， Korea （South）. IEEE ， 2020 ： 9724 - 9733 . doi: 10.1109/iccv.2019.00982 http://dx.doi.org/10.1109/iccv.2019.00982 Better to follow， follow to be better： towards precise supervision of feature super-resolution for small object detection WANG Z S ， WANG J Q ， LI Y L ， et al . Traffic sign recognition with lightweight two-stage model in complex scenes ［J］. IEEE Transactions on Intelligent Transportation Systems ， 2022 ， 23 （ 2 ）： 1121 - 1131 . doi: 10.1109/tits.2020.3020556 http://dx.doi.org/10.1109/tits.2020.3020556 Traffic sign recognition with lightweight two-stage model in complex scenes 10.37188/OPE.20233109.1366 TP394.1 A 1004-924X（2023）09-1366-13 李大湘,苏仲恒,刘颖.基于改进YOLOv4的道路交通标志识别[J].光学精密工程,2023,31(09):1366-1378. 李大湘,苏仲恒,刘颖.基于改进YOLOv4的道路交通标志识别[J].光学精密工程,2023,31(09):1366-1378. LI Daxiang,SU Zhongheng,LIU Ying.Road traffic sign recognition algorithm based on improved YOLOv4[J].Optics and Precision Engineering,2023,31(09):1366-1378. LI Daxiang,SU Zhongheng,LIU Ying.Road traffic sign recognition algorithm based on improved YOLOv4[J].Optics and Precision Engineering,2023,31(09):1366-1378. 国家自然科学基金资助项目（No.62071379）；陕西省自然科学基金资助项目（No.2017KW-013，No.2022GY-078，No.2022JM-374）；西安邮电大学研究生创新基金资助项目（No.YJGJ201902， No.CXJJLY202008） 2022-07-13 2022-08-08 2023-05-10 2023-05-15 37010460 光学精密工程 9 31