Research On SAR Image Target Recognition Based On Deep Learning

Synthetic Aperture Radar(SAR)is a kind of active coherent imaging radar,which has the characteristics of high-resolution,all-day,all-weather imaging and strong penetrability.The most important application of SAR in military field is to detect and identify specific military targets.Therefore,how to realize SAR Image Target Recognition is of great theoretical and practical significance.Recently,new Artificial Intelligence algorithms represented by Deep Learning theory have been successfully applied in the area of target recognition and classification,which has attracted wide attention at home and abroad.In this background,this thesis designed SAR Image Target Recognition algorithm based on Convolution Neural Networks(CNNs)and explored its performance.The main content is summarized as follows:(1)This thesis introduced and summarized the common algorithms of SAR Image Target Recognition and their advantages and disadvantages,followed by the analysis of key technologies and challenges faced by SAR Image Target Recognition based on Deep Learning.(2)Due to the problem of feature extraction and classifier design in SAR image target recognition is time-consuming and labor-intensive,a detailed study of CNNs in Deep Learning has been conducted,which summed up the key issues to improve CNNs performance.At the same time,an integrated SAR Image Target Recognition algorithm model based on the CNNs that can directly transfer the input image to the output category has been successfully constructed.On the basis of two inhibition measures of over fitting(L2 regularization technique and mechanism of Dropout),the algorithm model improved the fully connected layer structure of network and designed a SAR Target Recognition algorithm based on batch normalization operation,ReLU nonlinear activation function and AdaDelta gradient descent optimization algorithm.And the experimental results under the SOC and EOC acquisition conditions verified the active and effective learning ability of the proposed algorithm.(3)The thesis focused on panning sensitivity,attitude sensitivity and speckle noise and shading sensitivity derived from intensity sensitivity of SAR image in the practical application.A SAR Image Target Recognition algorithm based on Multi-sample Expansive CNNs was designed,which improved the learning ability,generalization ability and robustness of the algorithm.What's more,the sensitivity of the original design algorithm to sample size was discussed,and the adaptability of solving small-sample problems by SAR Image Target Recognition algorithm based on Multi-sample Expansive CNNs was studied.Experimental results showed that the proposed algorithm not only has a strong ability of learning features,but also has strong robustness to target translation,rotation,speckle noise and occlusion,and has a certain generalization ability under small sample conditions.