Research On Automatic SAR Vehicle Target Detection Algorithm Based On Deep Learning

SAR image target detection technology is widely used in modern military combat reconnaissance and civil terrain monitoring and other fields,and is the focus of research by scholars all over the world.With the development of neural network theory and computer technology,target detection algorithm based on deep learning has gradually become the mainstream of SAR image target detection tasks.However,the deep network usually has a large number of parameters,its operation requires a large amount of computing power support,and the running speed can not meet the real-time requirements,so the deep network detection algorithm is difficult to deploy on resource-constrained devices(such as mobile terminals).Considering the requirements of real-time performance and portability of the model,this paper improved the target detection algorithm Faster-RCNN in light weight,and designed a SAR vehicle target detection system based on the improved Faster-RCNN algorithm.The specific research contents are as follows:First of all,Article on the analysis of the characteristics of SAR image,on the basis of coherent spot of MSTAR image set suppression and image preprocessing operations,on the one hand,to optimize the quality of the test image,on the other hand the training set image data expansion,reduce network fitting risk,and with the aid of image information tagging software established the SAR image library vehicle target detection algorithm,target detection algorithm for SAR vehicle provides a comprehensive and effective image data.Lightweight improvement of Faster-RCNN algorithm.The improved Faster-RCNN algorithm is based on the framework of different parts of the model.In the first part,for the feature extraction network,the convolution structure of the feature extraction network is optimized by deep separable convolution,and the output receptive field of the feature extraction network is determined according to the image size and target size.The second part of the RPN web,first,combining with the clustering of SAR image target size generate anchor box dimension,and to improve the output of the RPN network pooling layer anchor frame quantitative way,retain more accurate anchor box position,reduce the calculation error on anchor box does not match,and then improve the maximum suppression algorithm is obtained by using the matrix iterative process.In the third part,global average pooling is adopted in the full connection layer of feature integration to reduce the redundancy of parameters in the full connection layer.Improved Faster-RCNN algorithm experiment.In this paper,the improved Faster-RCNN algorithm was trained and tested on the vehicle target detection image library of SAR images to detect the optimization effect of the improved Faster-RCNN algorithm on the mobile platform.Firstly,the influence of different improved methods on the model volume and detection accuracy was compared,and then compared with the single-stage target detection algorithm SSD.The experimental results show that the improved lightweight model can greatly reduce the model memory and algorithm computation while maintaining the original accuracy level,which can effectively meet the real-time requirements of SAR image target detection.Design and Implementation of Automatic SAR Vehicle Target Detection System.In this paper,Py Qt software is used to design and implement an automatic detection system for SAR vehicle targets based on the improved Faster-RCNN algorithm,and the system test image library is established to test the operation effect of the system.Through the test,the system can run smoothly and realize the detection function of SAR vehicle targets.