Research On Target Detection And Discrimination Based On Convolutional Neural Network For SAR Image Of Xidian University

Synthetic Aperture Radar(SAR)is a kind of high-resolution imaging radar that can work all day in all weather conditions and has widely been used in military,agricultural,forestry construction,geological exploration,and other fields.As the resolution of SAR image is getting higher and higher,how to fully mine the SAR image information has become a research focus of scholars in many countries.The detection and discrimination of targets in a SAR image are two key steps to interpret the SAR image.After the target detection and discrimination,we can remove clutter and obtain suspected targets.However,the performances of traditional target detection and discrimination algorithms are unsatisfactory in complex scenes and need to be improved significantly.In recent years,Convolutional Neural Network(CNN)has achieved great success in recognition,detection,segmentation,and tracking of targets in optical images.The powerful feature-extraction capability of CNN provides a new way for solving problems in the SAR image interpretation.This thesis focuses on the application of CNN in SAR the target detection and discrimination.The main accomplishments of this thesis are summarized as follows:1.The traditional two-parameter CFAR(Constant False Alarm Rate)algorithm is examined first.A super-pixel block approach is suggested to correct the target position because the detected target by using the two-parameter CFAR algorithm usually appears at the edge of the chip image.Furthermore,the traditional discrimination features of SAR targets are discussed and their shortcomings are analyzed through experiments2.The modifications on the common CNN structure are proposed from the following three perspectives to improve the discrimination accuracy of SAR targets and overcome the weaknesses associated with the traditional discrimination algorithms for SAR targets,including poor robustness in the discrimination features and difficulty in selecting the combination of features.First,it is modified from the perspective of fusing global and local features.The global and local features are extracted at the end of the network,and the two features are fused through the concatenating operation to achieve improving the accuracy of discrimination by the use of complementary features.Next,it is modified from the perspective of aggregating deep and shallow features.The features of the first four layers are aggregated,and jointly work with the features of the fully connected layer on the target discrimination to alleviate the problem of image information loss in the procedure of convolution and pooling operations.Lastly,it is modified from the perspective of deepening the network depth and effectively using the image space and channel information.The original convolution layer is replaced by a residual network module,and a Squeeze-and-Excitation variant module is added at the end of the network to solve the problem of weak network feature extraction capability.The proposed method is validated through experiments.3.In practical applications,the shortage of training samples and complicated scene are very common.Inspired by the transfer learning method,this thesis proposes a method by using the image style translation to expand the dataset and transferring model parameters trained on other data sets.The image style translation is performed on MSTAR dataset using the Cycle GAN technology,and the transferred dataset is used for the SAR target detection in Mini SAR dataset.Meanwhile,the parameters of the YOLOV3 model trained on the MSTAR dataset are transferred to the target detection in Mini SAR dataset to realize the model parameter transfer.Finally,the effectiveness of the method is verified via experiments.