Research Of Target Detection And Recognition In Large Scene SAR Image Via Deep Neural Network

An ATR system for SAR imagery(SAR ATR)is typically multi-staged to handle the SAR imagery in a divide-and-conquer approach.In order to recognize targets in SAR images,the traditional methods of SAR ATR(Automatic Target Recognition)consists of four independent main steps: detection,discrimination,feature extraction and recognition.Many factors between detection and recognition will greatly affect the recognition results,such as the difference of the size and the location of target between detection results and training samples.In order to quickly and effectively detect and recognize targets from the large scene SAR image,in this thesis,the proposed method can integrate the multi-staged SAR ATR as a whole system and directly recognize targets from large scene SAR image by encapsulating all computation in a single deep convolutional neural network(DCNN).The specific work is as follows:First of all,we summarize the development status of SAR target detection and recognition and deep learning in the field of image detection and recognition,and introduce the traditional SAR ATR key technology and basic deep learning algorithm theory in detail.Aiming at the shortcoming of hierarchical SAR ATR system,an end-toend method of target detection and recognition in SAR image based on deep network is proposed.Then,for the over-fitting problem caused by insufficient training samples for SAR images using deep learning method,we study the method of expanding the number of training samples based on Generative Adversarial Network(GAN).After introducing the basic theory of GAN,we attempt to use the Deep Convolutional Generative Adversarial Networks(DCGAN)to do the experiment of generating the SAR image for expanding the training sample set.Then we discuss the feasibility of GAN in the expansion of SAR image samples by several experiments.Finally,a method of cutting large scene SAR image to sub-images and a nonmaximum suppression algorithm for overlapping sub-images are proposed in this thesis.Then the framework of SAR D-ATR(Directly Automatic Target Recognition,D-ATR)is proposed.Several contrast experiments are designed,and experiments on the large scene SAR images with different background are carried out.The experimental results demonstrate that the D-ATR system shows high accuracy with a fast processing speed,and outperforms other methods.