Study On Target Recognition Method For Sequence High-resolution Radar Images Based On Deep Learning

High-resolution imaging radar mainly includes synthetic aperture radar(SAR)and inverse synthetic aperture radar(ISAR).Due to the advantages of day-and-night,all-weather,long range and high resolution,they have played significant roles in battlefield reconnaissance and situation awareness.Most of the traditional high-resolution radar target recognition methods are based on single image feature extraction and recognition.However,in ISAR,spotlight SAR,or circular SAR imaging,a high-resolution image sequence can be obtained by observing the same target continuously.In the image sequence,the target shape,structure,and backscatter intensity vary with the change of viewing angle and contain important features for target recognition.Therefore,how to exploit sequence image features has become a key issue to recognition performance improvement.Moreover,thanks to the characteristics of automatic feature learning,high accuracy,and strong generalization abililty,deep networks has been widely applied in the field of high-resolution radar target recognition.However,the existing deep learning-based high-resolution radar target recognition methods still suffer from: deficient of sequential information extraction and feature fusion,poor robustness to target unknown deformation,and network training difficulties under insufficient samples.Aiming at these issues,this dissertation takes in-depth analysis of the characteristics of high resolution radar images;then,it designs effective feature extraction and classification methods to achieve high accuracy and robust recognition of high resolution SAR/ISAR image sequences.The main contents of this dissertation are as follows:1.Aiming at the problem that the existing deep learning algorithms fail to utilize the characteristics of radar continuous imaging and only use single images for recognition,a SAR sequence target recognition method based on bidirectional convolutional recurrent network(BCRN)is proposed.Firstly,the spatial features of each frame in the sequence are extracted by a convolutional neural network(CNN)without fully connected layers;then,the sequence features are extracted by bidirectional long-short-term memory(Bi-LSTM).Finally,the recognition results are obtained by an averaged softmax classifier.The performance of the proposed method is evaluated on the SAR image dataset of ground vehicles,which shows robustness to large depression,configuration variants,and version variants.2.During sequence target recognition,the temporal features are difficult to be parallelly extracted,and the spatial information may be partially loosed.In view of this,a spatialtemporal ensemble convolution network(STEC-Net)is proposed.The spatial and temporal features of SAR sequential images are extracted simultaneously by causal and dilated 3-dimensional convolution,which are then hierarchically fused.The recognition results of SAR image dataset for ground vehicles show that the proposed network can reduce the training time consuming by 92% compared with traditional recurrent neural networks,while the recognition accuracy is increased by 1.11%.3.For ISAR sequential images,unknown target deformation exists due to the inherent imaging mechanism.To tackle the issue of effective deformation robust sequence feature extraction,a sequential adjustment ISAR network(SAISAR-Net)is proposed for ISAR sequence image recognition,where the affine transformation and deformable convolution are equipped to adjust the sequence image globally and locally.In addition,the lower convolutional network shares part of the parameters to avoid overfitting.The recognition results on the four-satellite ISAR simulation dataset show that compared with the existing deformation robust recognition methods,the proposed network improves the recognition accuracy by 7.33%,and it is robust to unknown deformations such as scaled,rotated,and combined transformation.4.The existing attention-based deep networks have weak performance on spatial,temporal,and channel features fusion,thus it is intricate to train them on high-resolution radar images.To tackle this issue,a hybrid Transformer ISAR sequence image target recognition method is proposed.Firstly,a spatial-temporal encoder based on the attention mechanism is constructed to extract the global and long-term robust features among sequential images.Then,a local feature encoder composed of 3D CNN is constructed to extract the local and short-term features.Finally,feature fusion and classification are performed by channel encoder-decoder and key-value query modules.Experimental results show that the hybrid Transformer provides inductive bias information for effective training of the attention mechanism and enhances local feature expression.Compared with the existing target recognition methods based on the attention mechanism,the recognition accuracy of the proposed method is 4.95% higher,and it is robust to the unknown deformation.5.For the challenges that the large-scale deep networks are difficult to be trained under limited high-resolution radar image samples,a deformation robust recognition method for few-shot sequential ISAR images based on semi-supervised transfer learning is proposed.Firstly,a sequential homography network is designed to extract robust sequence image features.Then,the semi-supervised and manifold hybrid losses are added to improve the generalization of the network.Finally,transfer learning of new classes with few samples is realized by optimal transfer mapping.Experimental results on 7-satellite simulation dataset show that the proposed method improves the average accuracy by 4.55% compared with the existing few-shot recognition methods,and it is robust to the unknown image deformation.