SAR Target Detection And Recognition Methods Based On Convolutional Neural Network And Knowledge Distillation

The rapid development of Synthetic Aperture Radar(SAR)technology makes SAR play an increasingly important role in military and civilian fields.As the key tasks in SAR image interpretation,target detection and target recognition are of great research significance.However,traditional methods of SAR target detection and recognition have been difficult to adapt to the increasingly complex and diverse SAR scenes.Since the convolutional neural network has a powerful ability of feature extraction,more and more researchers apply it to SAR target detection and recognition tasks,and have obtained better results than traditional methods.However,there still exist some problems in those methods based on convolutional neural networks: 1)Background clutter of SAR image is complex and diverse,and the information of background clutter is not fully utilized,which affects the detection performance of the network.2)The convolutional neural network usually contains lots of parameters and computations,thus it is difficult to deploy it to the resource-constrained radar devices.This thesis focuses on studying the above two problems through combining convolutional neural networks with knowledge distillation methods,and the main works are summarized as the following two aspects:1.Aiming at the problem that the information of complex and diverse background clutter is not fully utilized in the SAR target detection task,which affects the detection performance of the network,a SAR target detection method based on mutual learning Single Shot Multi Box Detector(SSD)is proposed.Inspired by the idea of mutual learning in knowledge distillation,two SSD networks with the same structure but different initializations are requested to assist each other in training.The classification loss function and regression loss function are designed for mutual learning,which helps each SSD network obtain additional supervision information from the other SSD network during training stage,including the confidence of target and background clutter,as well as the accuracy of target border positioning.The training strategy helps the network make full use of the information in clutters,better capture the difference between targets and clutters,and reduce false alarms and missing alarms.The experimental results based on Mini SAR dataset demonstrate that the proposed method helps the SSD network converge to a deeper and more robust local minima,and increases F1-score by 7.6 percentage points compared with the original SSD without changing the network structure and introducing additional parameters.What's more,our proposed method achieves the best experimental results in contrast to the similar SAR target detection networks.2.Aiming at the difficulty of deploying convolutional neural networks that contain lots of parameters and computations to the resource-constrained radar terminal devices,a SAR target recognition method based on knowledge distillation is proposed,whose purpose is to train a high-performance and lightweight SAR target recognition network.First,with the consideration of characteristics of SAR images,model size and performance,a convolutional neural network with channel-wise attention mechanism,called Channel-wise Attention Network(CA-Net),is devised for SAR target recognition.Then An attention-based channelwise network pruning method is presented to compress the network structure while keeping the performance of CA-Net as much as possible.To compensate for the performance penalty caused by the network pruning,a bridge connection based knowledge distillation algorithm is proposed to enhance the performance of the pruned network.The original CA-Net and the pruned CA-Net are regarded as the teacher network and the student network respectively,and the bridge connections are established between them to transfer the knowledge of the teacher network to the student network,which can improve the performance of the student network.The experimental results based on MSTAR dataset have proved the effectiveness and superiority of the proposed methods.Compared with other convolutional neural networks,the compressed CA-Net has fewer parameters and computations and achieves high performance.