Research On SAR Image Target Recognition And Classification Based On Deep Learning

Synthetic aperture radar(SAR)images technology is widely used in military reconnaissance,civil surveillance and other fields because of its strong anti-jamming ability.However,the traditional SAR images target detection technology can not be applied to all detection conditions since the feature extraction relies too much on the imaging configuration.With the help of more thorough data analysis,deep learning is undoubtedly the best way to solve this problem,meanwhile it has better generalization capability and generality.In view of the increasingly complex image data,how to obtain and summarize effective feature information based on the idea of deep learning,so as to realize a SAR image target recognition and classification system with more accuracy,stronger robustness,good real-time performance and easy transplantation to mobile devices is undoubtedly a new development space in the field of computer vision.Based on the detection algorithms technology proposed in the field,this paper focuses on the problems of reducing the complexity of network structure,multi-scale feature fusion and improving the accuracy of small target detection,and also puts forward some improved methods.The main research contents and contributions of this paper are as follows:Firstly,based on the imaging characteristics of SAR image and the influence of speckle noise,the appropriate target object data set is selected and preprocessed,such as denoising,slicing,position adjustment and data enhancement,so as to improve the recognition effect and enhance the data robustness;While retaining the original image label information,add manual label to it for subsequent positioning and recognition.Secondly,based on the optimization of feature fusion and small target detection,an improved YOLOv5 network is raised.A K-means++ clustering algorithm is proposed to calculate the a priori anchor box,disperses the initial clustering center,and improves the recall and accuracy of the model.While accelerating the training network fitting through the new activation function,a non-maximum suppression method based on confidence weighting is put forward to filter the preselected box to further improve the detection performance and reduce the false alarm rate of recognition.The transformer prediction head module is proposed to be added,which enhances the local perception ability and connects more the characterization feature information.It also speeds up the network calculation through the attention mechanism of the convolution module which can obtain the semantic information related to pixels,while the replacement of the weighted bidirectional feature pyramid and the addition of small target detection layer further retain the feature information of small targets,so that the fusion and splicing of shallow features and deep positioning information can be better,and the detection accuracy is improved.Thirdly,based on the idea of location sensitivity,an optimized Faster-RCNN network is put forward.By moving the convolutional layer of shared data forward and adding the position-sensitive score map and pooling layer,the partition evaluation determination and location regression are carried out from the feature scores of each part of the target object,which effectively improves the positioning and recognition performance of the network for the target,simplifies the network structure to a certain extent and improves the detection speed.Finally,a large number of simulations are carried out on the improved network,and compared with other mainstream excellent algorithm models in the same environment,which proves the correctness and feasibility of the optimization method,and also completes the expected improvement effect of SAR image target recognition and classification.It can not only meet the real-time requirements of most monitoring scenes,but also meet the requirements of mobility,portability and lightweight to a certain extent.