Chinese Version Light学术出版中心
Latest Issue
Issue 8 ， 2022
-
FPGA implementation of CCSDS compressor for snapshot mosaic hyperspectral images
Abstract：To solve the problem of large data volume and high transmission bandwidth requirement of snapshot mosaic hyperspectral images， the latest CCSDS 123.0-B-2 multispectral/hyperspectral lossless and near-lossless compression international standard is adopted to realize the FPGA-based lossless and near-lossless compression of snapshot mosaic hyperspectral images. By improving the BIP （Band-Interleaved Pixels） sequencing algorithm， a snapshot mosaic hyperspectral sample can be dynamically processed per clock cycle to 11 times higher of data throughput， and the problems of pipelining and parallelization difficulties and slow processing speed in the FPGA hardware implementation of snapshot mosaic hyperspectral image compression of this international standard have been effectively solved. The experimental results show that the overall logic resource usage of the implemented compressor is less than 10% after the layout and wiring of the FPGA on Xilinx XC7Z020CLG-4 platform， and the compression of a hyperspectral image can be completed in about 22 ms under 100 MHz system clock， with lossless compression performance between 2.66~4.30 bits/samples and the performance near-lossless compression is between 1.01~3.70 bits/samples， which can meet the application requirements of snapshot mosaic hyperspectral imaging technology in wireless handheld and unmanned airborne fields.Keywords：Hyperspectral image compression;snapshot mosaic hyperspectral imaging;CCSDS 123.0-B-2;FPGA implementation- 281
- |
- 42
- |
- 0
Published：2022-04-26 -
Three-dimensional cultural relic hole repair method combining structure and texture
Abstract：Due to external factors， the unearthed cultural relic model appears to be broken， and its virtual restoration is of great significance to archaeology. Most of the existing hole repair methods complete the hole based on the structure of 3D model， and the completed result is lack of color and texture information of the surface after the 3D（3-Dimension） model structure is repaired. Therefore， based on spatial structure and texture information of cultural relics， this paper proposes a 3D cultural relic hole repair method.First， in order to solve the problem of repairing the structure of 3D cultural relic， the algorithm based on the radial basis function is used to fill the holes of the mesh model ， and the vertices of the hole patches are adjusted by fitting the surface equation to better integrate with the original model. Second， for in order to complete the surface color and texture information of the 3D model， make the hole patch and the original model surface texture natural transition， the 3D problem is converted into 2D（2-Dimension） image inpaint problem， and a refined network is added into EdgeConnect to generate higher resolution result. Final， using Mudbox software to map the 2D image to the surface of the 3D model to fuse the results of structure and texture repair.The results show that the improved 2D inpaint network in this paper improves the performance of the evaluation metrics PSNR， SSIM and MAE by 0.54%， 0.217% and 6.52%， respectively. In addition， the combination of 3D structure repair and 2D texture repair makes the restoration of the cultural relic model more complete.The proposed method can effectively restore the mesh structure and texture information of the 3D model of terracotta army.Keywords：hole repair;deep learning;GAN network;texture inpainting;radial basis function- 182
- |
- 43
- |
- 0
Published：2022-04-26 -
Target operating range of dual band aerial remote camera
Abstract：In order to solve the problem of lack of unified evaluation standard and engineering closed-loop for target operating distance calculation model of aerial remote sensing camera in different bands， the research on target operating distance evaluation standard of dual band aerial remote sensing camera and engineering application case analysis were carried out. Firstly， according to the characteristics of aerial remote sensing， the three-dimensional projection transformation relationship of the target is proposed to obtain the equivalent observation size of the target under any observation condition. According to the different radiation imaging characteristics of visible and infrared dual band point target and area target， the target operating range calculation models are established respectively. Then， combined with an engineering application example， the results of different calculation models are analyzed and compared， and the calculation model suitable for long-range squint imaging of aerial remote sensing camera is given. Then， the information extracted from the actual flight image of an aerial remote sensing camera is used to reasonably select the signal to noise ratio（SNR） and modulation threshold in the calculation process of target operating distance. According to the modified threshold， the corresponding operating distance of large ships and tanks under different detection and recognition probabilities of visible light and infrared is analyzed. The analysis results show that when the flight height of the carrier is 18 km， the atmospheric visibility is 15 km and the visible light focal length of the aerial remote sensing camera is not less than 1.5 m， the detection distance of 141 km and 93 km for large ships on the sea and 105 km and 80 km for small targets such as land tanks can be realized with 50% probability. Under the condition that the infrared focal length is not less than 1m， the detection distance of 203 km and the recognition distance of 140 km for large ships on the sea and 44 km and 37 km for small targets such as land tanks can be realized with a 50% probability. The above simulation results are consistent with the reality， and the results of this paper can be better applied to engineering practice.Keywords：aerial camera;visible and infrared;operating distance;detection and identification;SNR;modulation;Contrast- 133
- |
- 36
- |
- 0
Published：2022-04-26 -
Cloud classification based on ensemble learning combining with deep neural network and FSVM
Abstract：Accurate cloud classification is of great significance for meteorological monitoring. Traditional machine learning models rely on hand-craft featurs， which is sensitive to noise data and the generalization ability is also poor. Deep neural network can automatically learn the depth features of image， but it is not good at image edge and detail classification， this paper studies on the basis of the above problems. First， the spectral features and texture features are extracted from himawari-8 satellite images to train fuzzy support vector machine （FSVM） model. At the same time， different channels of cloud images are selected to train deep neural network to learn the depth features for cloud classification. Finally， according to the characteristics of different models， the output of the two models is fused by ensemble learning to improve the classification accuracy. This article designs a cloud classification model based on ensemble learning which fuses deep neural network and FSVM. It combines the advantages of different models and makes use of the complementarity between different models to improve the robustness and reliability of the model.The experimental results show that： compared with model which uses a single model alone， the ensemble learning method proposed in this article has better performance in different evaluation indicators， The average POD， FAR and CSI were 0.9245， 0.0796 and 0.8581 respectively； this method also has better recognition effect when compared with other cloud classification models； in specific cases， it is found that this method has higher recognition accuracy in clouds mixed regions， and it can identify cloud edge and cloud details more accurately.This model can satisfy the requirements of stability， reliability， high precision and strong generalization performance of cloud classification model.Keywords：Himawari-8;cloud classification;deep neural network;FSVM;ensemble learning- 150
- |
- 30
- |
- 0
Published：2022-04-26 -
Multi-feature fusion high-throughput dPCR fluorescence image recognition
Abstract：The results of traditional high-throughput dPCR fluorescence image analysis are prone to low positive spot recognition rate due to false positive points and non-specific amplification. Therefore， in this paper， a multi-feature fusion high-throughput dPCR fluorescence image recognition method （HDFINet） is proposed to improve the accuracy of high-throughput dPCR fluorescence image recognition. Firstly， a up-bottom structure is introduced in the feature fusion part so that the lower layer features can be used more effectively in the top layer. In the up-bottom structure， channel attention is used to assign channel weight of fluorescent image， and spatial attention is used to assign corresponding weight of fluorescent image pixels in the feature map， so that the feature map can better respond to the feature of fluorescent image positive points. Then， the confidence of the bounding box of positive points was calculated by using the adaptive Intersection-over-Union （IOU） in RPN to reduce the possibility of loss of positive points information. Finally， ROI Align re-fixed the size of the features in the candidate areas of fluorescent images， and then input them to the full connection layer and fully convolution layer to perform category and regression box regression and output positive point recognition results. The experimental results show that the HDFINet network proposed in this paper has a high recognition rate and can effectively realize the recognition of positive points in fluorescent images. Compared with YOLOv4， VF-Net， and GROIE， the comprehensive index F1 of the method in this paper is the highest， compared with the classic deep learning Network Mask R-CNN network， this method increases the true positive rate of positive points by 1.13%， the positive predictive value by 0.36%， and the value of the comprehensive index F1 by 0.75%. The HDFINET network proposed in this paper has good recognition performance and can effectively identify positive spots in fluorescence images， which has reference value for other fluorescence image analysis and research.Keywords：dPCR;deep learning;fluorescence image;positive points- 132
- |
- 21
- |
- 0
Published：2022-04-26
Modern Applied Optics
-
Design on constant coherent light intensity fringe locking system for scanning beam interference lithography
Abstract：In order to improve the processing efficiency and processing accuracy of scanning beam interference lithography machines， expand their processing range， and reduce maintenance costs， a scanning beam interference lithography system is proposed， which can realize the synchronization and real-time adjustment and control of the interference fringe period， direction， and phase. Additionally， the design of the most critical constant coherent light intensity interference fringe phase locking system is emphatically introduced. First， the principle scheme of the scanning beam interference lithography tool with real-time adjustment and control of the interference fringe period， direction， and phase in lithography is introduced， and the constant coherent light intensity heterodyne acousto-optic frequency-shifted interference fringe phase locking control system is designed and implemented. Then， the theoretical model of the control system is analyzed， the actual model of the system is identified， high-order linear fitting is performed， and the system controller is designed. Finally， debugging and fringe lock control are performed on the basis of this controller. The experimental results show that after fringe locking control is turned on， low-frequency disturbance components below 100 Hz are significantly suppressed， and the residual error of the interference pattern phase locking is 0.0693 rad （3σ）； that is， the phase change is less than a factor of ±0.01 of the interference fringe period. Additionally， when the beam offset adjustment error reaches 100 μm， the interfering fringe lockability remains stable. The system can not only achieve high-precision locking of interference fringes under different interference attitudes， but it also has a large beam control error margin， which meets the needs of a scanning beam interference lithography machine with real-time control of the interference fringe period， direction， and phase.Keywords：scanning beam interference lithography;fringe locking;heterodyne phase measurement interferometer- 140
- |
- 21
- |
- 0
Published：2022-04-26 -
High-precision thermal deformation measurement of optical remote sensor optical-mechanical structure
Abstract：In the complex space orbit thermal environment， the optical structure of an optical remote sensor is thermally deformed. Thermal deformation affects detection accuracy. To solve the problem of thermal deformation measurement of the optical-mechanical structure of optical remote sensors， an integrated monitoring method is proposed based on digital photogrammetry and fiber Bragg grating （FBG） thermal deformation. A fiber grating layout method and a deformation measurement model algorithm with a thermal decoupling function are developed. The relationship between the spectral variation of the FBG and strain and temperature is analyzed. A digital photogrammetry experimental system is built to measure the strain and temperature fields of the optical machine structure. The FBG sensing network enables high-precision measurement of the thermal strain of optical machine structures. The results reveal that the integrated system can realize the simultaneous measurement of the displacement field， strain field， and temperature field of an optical machine structure. The accuracy of the structural deformation measurement after thermal decoupling reaches 0.02 mm， which can be used to reasonably predict the thermal deformation of optical machine structures on orbit. The proposed measurement method can potential be applied to the on-orbit monitoring of optical machine structures of space optical remote sensors.Keywords：optical remote sensor;thermal deformation;digital photogrammetry;FBG strain sensor;FBG temperature sensor;experimental verification- 169
- |
- 46
- |
- 0
Published：2022-04-26
Micro/Nano Technology and Fine Mechanics
-
Spatial-spectral separable convolutional neural network for cell classification of hyperspectral microscopic images
Abstract：In recent years， with the development of computer science，deep learning plays a critical role in the classification of hyperspectral bloodcell images. However， traditional deep learning models require a large amounts of manually annotated training data， and ignore the nature of “graph-spectral uniformity” property of hyperspectral image. As a result， these methods can not explore the intrinsic information of hyperspectral images. In addition， traditional convolutional neural network methods have too many parameters， which takes a great deal of time to be trained. Aiming at these two shortcomings， a spatial-spectral separable convolutional neural network （S3CNN） is proposed to improve the classification performance of bloodcell hyperspectral image and reduce the complexity of the model.First， due to the spatial consistency of the hyperspectral bloodcell image distribution， a spatial-spectral combined distance （SSCD） was proposed to select the spatial-spectral nearest neighbor of each pixel and expand the training samples. At the same time， in the following neural network model， a group of depth convolution and point convolution are used to replace classical convolution and optimize the complexity of the model.The experimental result on bloodcell1-3 and bloodcell2-2 datasets show that the overall classification accuracies reaches 87.32% and 89.02%， respectively. Compared with other classification algorithms of bloodcells， the proposed S3CNN achieves much higher classification accuracy. The training time of the separable convolution model is 27% less than that of the classical convolution model.Experimental results show that the proposed S3CNN is an effective method to improve the classification performance of hyperspectral bloodcell and reduce model training time.Keywords：hyperspectral image;bloodcell classification;convolutional neural networks;spatial-spectral combined distance;separable convolution- 170
- |
- 215
- |
- 0
Published：2022-04-26 -
Colorectal polyp segmentation algorithm using DoubleUNet network
Abstract：Colorectal polyps are different in size， color and texture， and the boundaries between the polyps and the surrounding mucosa are not clear， leading to significant challenges in polyp segmentation. In order to improve the segmentation accuracy of colorectal polyps， this paper proposes an improved DoubleUNet network segmentation algorithm. The algorithm first de-reflects the polyp image， and the training dataset is amplified by data-augmentation method； then introduces an attention mechanism in the decoder part of the DoubleUNet network， and replaces the atrous spatial pyramid pooling module of the network with a densely connected atrous spatial pyramid pooling module to improve the ability of the network to extract features； finally， in order to improve the segmentation accuracy of small targets， the Focal Tversky Loss function is proposed as the loss function of this algorithm. The accuracies of the algorithm in the Kvasir-SEG， CVC-ClinicDB， ETIS-Larib， ISIC， and DSB dataset are 0.953 0， 0.964 2， 0.815 7， 0.950 3， 0.964 1， respectively， while the accuracies of the DoubleUNet algorithm in the above datasets are 0.939 4， 0.959 2， 0.800 7， 0.945 9， 0.949 6. The experimental results show that the algorithm in this paper has a better segmentation effect than the DoubleUNet algorithm， which can effectively assist physicians to remove abnormal tissues of colorectum and thus reduce the probability of cancerous polyps and it can be applied to other medical image segmentation tasks as well.Keywords：image segmentation;colorectal polyps;dilated convolution;attention mechanism;conditional random field- 168
- |
- 43
- |
- 0
Published：2022-04-26 -
Point cloud registration using neighborhood point information description and matching
Abstract：Point cloud registration is an important part of reverse engineering， machine vision and other technologies in modern manufacturing. Its efficiency and accuracy have an important impact on the acquisition of product data model. In order to improve the accuracy and efficiency of 3D object point cloud registration， a point cloud registration method using Neighborhood point information Description and Matching （NDM） is proposed. Firstly， under three radius ratios， feature points are extracted according to the change of curvature， measurement angle and eigenvalue property； secondly， the improved normal vector angle， point density and curvature are calculated to obtain multi-scale matrix descriptor； then， a k-dimensional tree is established for descriptors， and the matching relationship is preliminarily established. The combination of geometric feature constraint and rigid distance constraint is proposed to eliminate the wrong points； finally， the k-tree improved iterative closest point （ICP） algorithm is used to complete the accurate registration. In this paper， two groups of experiments are designed， which are real object point cloud registration and Stanford model simulation real object registration. The results show that the algorithm solves the limitations of the classical ICP， and improving the registration accuracy by 2-5 orders of magnitude. Compared with other algorithms， the registration accuracy of this algorithm is improved by at least 29%， and the efficiency is increased by 58%； in the Stanford database simulation experiment， the registration accuracy is improved by 1%-99%， and the registration time is reduced by 3%-94%. It is proved that this algorithm is an effective registration method of point cloud on object surface， which can improve the registration accuracy and efficiency， and has good robustness.Keywords：reverse engineering;machine vision;point cloud registration;neighborhood point information;iteration closest point- 181
- |
- 36
- |
- 2
Published：2022-04-26 -
Lightweight small object detection network with attention mechanism
Abstract：In order to improve the detection ability of small objects in complex backgrounds， such as drone aerial images， this paper proposes a lightweight small object detection network that introduces an attention mechanism based on the YOLOv4 network. First， a multi-scale fusion module is added to the channel attention mechanism and construct a multi-method feature extractor， and then embed the designed channel attention module into the YOLOv4 feature extraction network to enhance the network's ability to focus on the region of interest in the image； then improve YOLOv4 network structure， increase the fusion of shallow feature layer and deep feature information to obtain rich resolution information； finally adopt channel pruning and knowledge distillation strategy to optimize the model of the improved network， and greatly reduce it under the premise of small accuracy loss the number of model parameters. The experimental results show that in the drone aerial photography data， the lightweight small object detection network proposed in this paper reduces the model size of the original network by 93.6%， the reasoning speed increases by 52.6%， and the mAP increases by 2.9%； in the cloth defect dataset， the model size is reduced by 92.1%， the reasoning speed is increased by 49.5%， and the mAP is increased by 2.2%， effectively improveing the detection of small objects in complex backgrounds and realizeing a lightweight network.Keywords：object detection;YOLOv4 network;attention mechanism;feature fusion;network compression- 312
- |
- 72
- |
- 2
Published：2022-04-26 -
GNSS-R significant wave height inversion method based on delay-Doppler maps
Abstract：In order to realize the detection of significant wave height， this paper uses GNSS-R （Global Navigation Satellite System Reflection） to propose an significant wave height inversion method based on delay-Doppler maps for signal-to-noise ratio（SNR） calculation. Compared with active detection such as traditional radar， GNSS-R is a passive observation using forward scattering of satellite signals， without transmitter. It has the advantages of more signal sources， low power consumption， low cost and light weight， which is convenient to be built on space-borne， airborne and ship-borne platforms. Due to the different significant wave height， the reflected signal received by the receiver is also different. Through the analysis of the reflected signal， we can get the physical characteristic information of different reflectors， so as to reflect the required significant wave height information. In this paper， the delay-Doppler map is generated by the reflected signal， and the corresponding SNR definition is given to obtain the significant wave height. In order to verify the feasibility of the algorithm proposed in this paper， the inversion experiment of significant wave height on board is carried out， and the performance is compared and verified with the traditional estimation method using interference complex field （ICF） for inversion. The results show that the root mean square error of the proposed algorithm is 0.098 m， the R2 value is 77 %， and the correlation coefficient is 0.82， which are better than those of the ICF method. The proposed algorithm has better consistency with the significant wave height data of the wave radar， which improves the accuracy of the significant wave height estimation based on GNSS-R technology， and can provide services for the significant state information perception of ship navigation.Keywords：GNSS-R marine remote sensing;significant wave height;signal-to-noise ratio;delay-Doppler map- 119
- |
- 15
- |
- 0
Published：2022-04-26
Information Sciences
Loading...
0