Research On Target Recognition Method With Synthetic Aperture Radar Image

Synthetic aperture radar(SAR)is a high-resolution imaging radar working in the microwave range,which has played a key role in the field of national defense and security,environmental monitoring,and energy detection on the account of its work capability of day-and-night and all-weather.The purpose of SAR image interpretation is to extract useful information from complex ground targets based on the physical attributes of the target and background according to different task requirements,so as to realize the transformation from image to intelligence information.Automatic target recognition(ATR)is dedicated to automatically detecting targets of interest from large target scenes and realizing target identity reasoning,which is one of the key technologies to realize intelligent SAR image interpretation.Recently,SAR ATR technology has gained remarkable advancement in the both theoretical methods research and system design along with machine learning and deep learning technology development.However,due to the limited data acquirement capability of senor,expensive annotation cost,and the unique imaging mechanism of SAR sensor in the real scenarios,it gives rise to significant challenges for training data set establishment,robust features extraction,and high-accuracy classification model design in SAR ATR tasks.This dissertation focuses on the key problems including sample annotation,target feature extraction,classification criterion,and so on in SAR ATR tasks,and carries out theoretical analysis,method research,and experimental verification with measured SAR data from the aspects of deep feature extraction,multi-view learning,weakly supervised learning,and few-shot learning.The main research contents are as follows:(1)With regard to SAR target recognition problems under complex extended operation conditions(EOCs),an extended convolutional capsule network is proposed.The proposed method first uses multiple dilated convolution layers to extract multi-scale features.Then,the idea of attention mechanism is introduced into multi-scale channels layer to adaptively highlight and suppress features.To mine effectively discrimination information from feature spatial structure,a capsule unit-based feature pose preserving layer is designed to learn the spatial structure relation among features,which can effectively mine the potential information from SAR image,thereby improving target recognition performance.(2)Focusing on the problem of unsatisfactory recognition performance because of limited information in a single-view SAR image,a supervised sparse representation multiview recognition method is developed.In the training phase,the proposed method makes full use of the supervision information provided by the training sample labels to learn the representation dictionary and classifier simultaneously,and leverages classification error to adjust the dictionary learning procedure.In the testing phase,the azimuth interval constraint in the multi-view joint optimization model is relaxed,while broadening the scope of application of the model,which can extract the common and complementary characteristics among multi-view SAR images,thereby promoting target recognition performance.(3)Aiming at an expensive annotation cost problem for SAR image in ATR tasks,a deep learning-based weakly supervised learning method is proposed.Considering that the target recognition model may prone to overfit in the case of the limited annotated samples,Bayesian inference is introduced into the processing of parameter optimization of convolutional neural network.To use efficiently a great volume of unlabeled samples to promote recognition performance,a sample selection strategy for weakly supervised classification and recognition tasks is proposed to complete the selection and labeling of unlabeled samples.Finally,the classification model is trained repeatedly with labeled samples.By so doing this,while reducing the cost of sample labeling,the recognition performance of the model is improved.(4)Focusing on the SAR target recognition problem in some scenes with very scarce training samples,a meta-Siamese subspace classification network is proposed.The proposed method integrates the ideas of meta-learning,metric learning,and subspace learning into one unified framework.The proposed method extracts discriminative embedding features by designing the feature embedding model with Siamese network structure,and designs a classifier by leveraging the idea of subspace learning to complete target identity reasoning.Also,the idea of “learn to learn”from meta-learning is introduced to the process of model training,which can realize stable SAR target recognition where training samples are extremely scarce.