Research On Detection And Recognition Of Weak And Small Targets For SAR Images In Complex Scenes

Synthetic Aperture Radar(SAR)is an active remote sensing payload that transmits and receives electromagnetic waves for all-day,all-weather,wide-area surface information acquisition,and is therefore widely used in many fields such as agricultural census,disaster monitoring and battlefield reconnaissance.The intensity of the target in SAR images reflects the strength of its radar echo signal and characterizes the backward scattering characteristics of the target,which provides an effective source of information and data for SAR automatic target recognition.However,the unique imaging mechanism and image characteristics of SAR constrain the performance and generalizability of SAR target detection and recognition algorithms,especially the detection and recognition of weak and small targets in large scenes.In this dissertation,the research takes the practical demand of SAR weak and small target detection and recognition in large scenarios as the guide,based on the data characteristics of SAR and the methodological characteristics of deep neural networks,and focuses on the problems of weak and small target detection and recognition such as small scale,weak response and small sample size of targets in SAR automatic target recognition.The research work of this dissertation is summarized as follows:(1).A small-scale SAR target detection method based on the revised Bhattacharya distance is proposed to address the difficult problem that smallscale SAR targets are easy to miss detection in complex scenes.Firstly,to address the challenge that small-scale targets with little pixel information are easy to miss detection,the semantic information of deep features is aggregated using the feature pyramid based on content-aware reassembly of features to enhance the discriminability of shallow features.Secondly,for the problem that small-scale target localization is demanding and the existing methods are sensitive to position deviation,the revised Bhattacharya distance is proposed to accurately measure the bounding box relationship.Finally,a multi-stage cascade detector is constructed to optimize the whole model step by step for the problems of low accuracy of small-scale target localization and difficulty of model optimization.The experimental results show that the proposed method significantly improves the detection of small-scale targets on the LS-SSDD-v1.0 dataset,with a recall of 84.4% and an average precision of 76.6% under an intersection ratio threshold of 0.5.(2).Aiming at the problem that SAR target response is weak in complex scenes and prone to missed detection and false alarm,a weak-response SAR target detection method based on the hierarchical guidance of local coefficient of variation saliency map is proposed.Specifically,for the problem that the weakresponse target is not conspicuous in the image,the coefficient of variation is extended to suppress background clutter in the image,and the center-surround difference algorithm is improved to generate the saliency map.A weak-response SAR target detector with hierarchical extraction and fusion of SAR images and saliency maps is constructed for the problem that weak-response targets are highly similar to some backgrounds and the model is difficult to converge.The experimental results show that the proposed method effectively improves the detection performance of weak targets,with precision and recall of 96.8% and96.0% on the mini SAR dataset,and 85.7% and 85.2% on the FARADSAR dataset,respectively.(3).For the problem of limited target samples in SAR images in complex scenes,a fast recognition method based on feature metrics for small-sample SAR targets and a highly plastic incremental recognition method based on knowledge inheritance for small-sample SAR targets are proposed.Firstly,a multi-task constrained deep metric network model is constructed for the problem of incomplete data of small-sample targets,which solves the problems of dependence on support samples,poor robustness and time-consuming prediction of existing metric learning-based methods.Secondly,to address the problem of small-sample targets without typical exemplars,a herding algorithm is used to select and manage target exemplars to alleviate the problem of catastrophic forgetting.Finally,to address the problem of difficult model expansion for smallsample targets in incremental recognition scenarios,a knowledge inheritance module is proposed for migration of old model recognition capability,which significantly improves the model stability,and the use of unstructured pruning to initialize new model parameters effectively enhances the model self-adaptation.The proposed small-sample fast recognition method achieves 83.26%recognition accuracy with only 20 training samples per class of target and takes1.16 s to predict.The accuracy of the proposed small-sample incremental recognition method is 93.6%,94.1% and 95.1% for the three incremental recognition scenarios,and the training time for the last task is 61.7s,43.7s and63.1s,respectively.