Mutil-aspect SAR Target Recognition Based On Deep Learning

Different from the optical imaging system,Synthetic Aperture Radar(SAR)imaging system collects the information of a ground object by sending and receiving electromagnetic wave.And the signal will be translated to SAR image by digital signal processing methods.SAR will not be affected by light or weather condition,and it can penetrate the shelter to a certain degree.Thus,SAR images can provide a lot of valuable information.However,the inherent speckle noise,geometric distortion and high azimuth sensitivity of SAR images bring a great difficulty to image interpretation,which makes the SAR automatic target recognition method with high research value.With the improvement of SAR technology,the data collecting modes become more flexible.Circular SAR and geosynchronous SAR,bring new direction and challenge to the SAR imagery interpretation.Compared with the single aspect SAR image,the multi-aspect images contains the space-varying scattering features of the target in addition to the basic backscattering features.Starting from the regular SAR target recognition based on a single aspect,with the possibility of acquiring multi-aspect SAR images the article further studied the SAR target recognition method based on multi-aspect SAR image sequences.The main work is as follows:Firstly,based on the understanding of the traditional SAR target recognition methods and the inherent characteristic of SAR images,the difficulties and some existing problems of the traditional recognition methods are analyzed.And in order to solve these problems,the deep learning methods are introduced.Secondly,for the traditional SAR recognition task,the residual neural network(ResNet)is used to learn the high-level abstract feature of SAR targets,and Dropout layer is introduced into the residual building block to alleviate the overfitting problem caused by small sample data.In addition,in order to make the features more discriminative,the center loss function which is used to learn the intra-class compactness is adopted and combined with the Softmax loss function to supervise the training of the network.Thus,the generalization and the recognition performance of the SAR target recognition method under the small sample data set can be improved.Finally,for the multi-aspect SAR images,based on the single aspect SAR target feature extraction model,a recurrent neural network model is introduced,and then a multi-aspect SAR target recognition method based on recurrent learning is proposed,which utilizes the high-level abstract feature extraction ability of the ResNet and the context features learning ability of the bidirectional long-term memory network(BiLSTM).In addition to learn the features obtained from the single aspect SAR image,the combined network can continue to learn the space-varying features from multi-aspect,which incorporates more characteristics of multi-aspect SAR targets and can improve the recognition accuracy as well as the generalization ability of the method.