Abstract:To improve the accuracy and real-time performance of infrared (IR) dimsmall target detection, an IR dimsmall object detection algorithm based on an improved multi-scale fractal feature was presented.Computational analysis of the multi-scale fractal feature related to the fractal parameter K (MFFK), which was used for IR image enhancement in the algorithm, was performed. First, an improved multi-scalefractal feature (IMFFK) was presented to perform image enhancement after substituting the equation for computing fractal dimension into the equation for computing MFFK using the covering-blanket method. Thereafter, a computationally efficient IR dimsmall target detection algorithm was presented, in which the computation of IMFFK was simplified and an adaptive threshold was used to segment targets of interest from the background. Finally, the effect of primary parameters on image enhancement and computational cost was analyzed based on the simulation images. The IR real-world images were subsequently used to evaluate the detection performance of the proposed algorithm, and comparisons with state-of-the-art detection algorithms based on local contrast measureare performed. The proposed algorithm was capable of simultaneously detecting dimsmall and large targets in an IR image, irrespective of whether the targets were bright or dark, even though false alarms were detected in some scenarios. It is also capable of reachingapproximately 30 frames per second for low-resolution IR images (320×240). The proposed algorithm exhibitssatisfactory applicability and can be used to detect targets with high local contrast in an image.
Abstract:Ship detection from optical remote sensing images is easily disturbed by complex backgrounds, such as clouds, islands, and sea clutter. In this paper, a novel ship detection method was proposed to solve these problems. First, to solve the change in target size, multi-scale saliency maps were generated using a spectral residual visual saliency model, and the optimal saliency map was adaptively selected using the Gini index. Further, considering the problem of missing detection caused by the threshold segmentation, a two-stage segmentation method was proposed to separate the target and background pixels. The local maximums of the saliency map were then obtained using image expansion, and the k-means algorithm was adopted to determine whether each local maximum belongs to the target pixel or background pixel. The accurate candidate locations were obtained using the local threshold segmentation. Finally, the rotation invariant feature based on the radial gradient transform was introduced to further eliminate false alarm. The experimental results show that the proposed detection method can successfully detect ship targets of different sizes and directions and effectively overcome complex background interference. Additionally, the detection accuracy is 93%, and the false alarm rate is 4%, which are better than other saliency-based ship detection methods.
Abstract:Object detection, which is a fundamental visual recognition problem in computer vision, has been extensively studied in the past few decades and has become one of the popular research areas in the world. The aim of object detection is to accurately locate specific objects in a given image and assign a corresponding label to each object. In recent years, Deep Convolutional Neural Networks (DCNN) have been used in a series of developments in object detection and image classification owing to their powerful capabilities of feature learning and transfer learning.It has garnered considerable attention in the field of computer vision for object detection. Therefore, the method of applying CNN in target detection to obtain better performance is an important topic for research.First, we reviewed and introduced several types of classic object detection algorithms.Next, we considered the generation process of the deep learning algorithm as a starting point, analyzed the technical ideas and key problems of DCNN in the application of target detection, and provided a comprehensive overview of various target detection methods in a systematic manner. Finally, in view of the major challenges in target detection and deep learning algorithms, we provided future development scope and direction to promote the study of target detection using deep learning.
Abstract:The Tiny YOLOV3 target detection algorithm has a high error rate for small targets, such as pedestrians, in real-time detection.Therefore, this study aimed to improve the feature extraction network, prediction network, and loss function of the algorithm.First, a two-step convolution layer was added to the feature extraction network to replace the maximum pooling layer in the original network for downsampling. Second, the traditional convolution was replaced with an anti-residual block constructed by a deep convolutional convolution to reduce the model size as well as number of parameters and increase the high-dimensional feature extraction.Third, based on the original two-scale prediction of the network, a scale was added to form a three-scale prediction. Finally, the boundary box position error in the loss function was optimized.The experimental results demonstrat that the improved Tiny YOLOV3 algorithm achieve a target detection accuracy that IS 9.8% higher than the original algorithm, satisfied the real-time requirement, and demonstratedrobustness.The proposed method can better extract target features, and the multi-scale prediction and improvement of the boundary box position error can detect targets more accurately.
Keywords:target detection;Tiny You Only Look Once V3(YOLOV3);depth separable convolution;anti-residual block;multi-scale prediction
Abstract:To solve the problem of detecting small infrared targets ininfrared optical systems under complex backgrounds, an information processing model based on visual feature integration was established, and a dim small target detection method based on visual feature integration was proposed. First, the difference of Gaussians model of the receptive fields of retinal ganglion cells was used for processing the primary information of infrared images and initially detecting dim small targets. Then, the features contai
Keywords:computer vision;target detection;dim small target;visual feature integration
Abstract:In the three-dimensional (3D) precision measurement of large component, the detection accuracy of cooperative targets is low due to complex structure of large components and various measurement environment. To solve this problem, a multi-type cooperative target detection method using improved YOLOv2 convolutional neural network was proposed. Firstly, the data augmentation method combined with WGAN-GP was employed to amplify the number of cooperative target images. Secondly, the convolutional layer dense con
Abstract:Target detection for hyperspectral image(HSI) is a hot topic, due to its important military and civilian applications. This paper proposes a novel target detection algorithm for HSI based on tensor representation. The algorithm employs tensor analysis including CP decomposition and tensor block decomposition to implement blind source seperation to the hyperspectral data. Effective spatial and spectral features of blocks of local image were extracted. And then the algorithm establishes a detection model based on sparse representation and collaborative representation. Experiments were conducted to evaluate the performance of our approach under multiple scenes with complex background. From the visual representation of the results, it can be concluded that the proposed approach effectively extracts the spatial-spectral features under scenes with strong noise and complex background. The approach has good ability to suppress the background and the target is salient. In addition, the performance of the approach is evaluated by quatitative metrics such as receiver operating curve (ROC) and area under the ROC curve (AUC). Taking the popular HSI image-Sandiego image as an example, the approach achieves 90% detection rate with the false alarm rate of 10% and the AUC is greate than 0.95. Our approach ourperforms the other popular ones.
Abstract:Target detection for Hyperspectral Images (HSIs) is gaining importance owing to its important military and civilian applications. This study proposed a novel target detection algorithm for HSIs based on tensor representation. The algorithm employed tensor analysis including CP and tensor block decompositions to implement blind source separation on hyperspectral data. First, effective spatial and spectral features of the blocks of local images were extracted. Then, a detection model based on sparse and colla
Abstract:To improve the performance of the clutter suppression and small target detection, a detection algorithm for a small target in sea clutter was proposed, based on the spatio-temporal chaos analysis. First, the sea clutter phase space was reconstructed as a chaotic dynamical system, and the chaotic parameters of the sea clutter sequence image were extracted to verify that the sea clutter owns chaotic properties in the spatial and temporal domains. Furthermore, the spatial chaotic reconstruction function, the t
PAN Rong,SUN WeiVol. 25, Issue 10s, Pages: 221-227(2017) DOI: 10.3788/OPE.20172513.0221
Abstract:Aiming at the problem of difficult small target detection in high resolution images, combined with region-of-interest (ROI) extraction strategy in target detection method based on candidate region and regression strategy in target detection algorithm based on regression, deep learning target detection algorithm based on pre-segmentation and regression (Quad-ssd) was proposed. As fast-RCNN series implement image location and classification separately, small targets could be detected but detection time was too long. YOLO series method used regression method to implement classification and location for targets in images at the same time. As only high-level features were used, detection accuracy for small target was not enough. Therefore, quad tree was used to extract interest target of original images, and target detection method based on regression was used to implement detailed relocation and classification for targets in interested region. Compared with traditional Fast-RCNN method and deep learning method based on regression of YOLO series, target detection algorithm of deep learning based on quad tree has obvious advantages in accuracy and speed. The experimental results show that compared with Fast-RCNN, accuracy of Quad-ssd algorithm is improved by 6.5% and reaches 74.9% at the time of target detection. The detection speed is improved greatly; reaching 45 f/s, and can satisfyrequirements of timeliness.