Study On Key Techniques For SAR Target Recognition Based On Deep Learning

Synthetic Aperture Radar(SAR)is a wideband imaging radar featured by all-weather,all-day,and long operational range capabilities.SAR breaks through the limitation of single aperture of traditional radar and achieves high-resolution imaging through synthesis of virtual synthetic apertures,providing more information about the target shape and scattering characteristics than narrowband radar.SAR automatic target recognition(ATR),especially typical military target recognition,plays an important role in battlefield reconnaissance,situational awareness,and detection guidance.Different from optical images,the gray-scale single-polarized SAR image has strong anisotropy and is easily interfered by clutter,adding difficulties in ATR.Traditional SAR ATR algorithms require manual feature extraction of SAR images,which relies heavily on the prior knowledge of experts.In addition,the steps of feature extraction are tedious and are not conducive to the intelligence and automation of SAR ATR methods.In recent years,deep learning has developed rapidly and found wide applications in SAR ATR tasks thanks to its strong automatic feature extraction capability.Although such methods do not require a great deal of expertise,they have certain disadvantages such as poor separability of learned features,redundancy of learned features,and decrease of recognition performance with insufficient training samples.To solve the above problems,this dissertation studies effective deep network structures and increases the recognition accuracy of SAR ATR.The main research contents are as follows: 1.First,the moving and stationary target acquisition and recognition(MSTAR)dataset and a typical deep learning structure,i.e.the Convolutional Neural Network(CNN)are introduced.The principles,basic structure and training methods are also discussed.Then,in order to reduce the negative effect of the highly correlated clutter in the MSTAR dataset in recognition,a morphology-based SAR image segmentation algorithm is proposed to isolate the target region.Finally,the accuracy of the designed network structure under different observation conditions is given to prove its effectiveness.2.A novel structure,namely the large-margin softmax batch-normalization CNN(LM-BN-CNN)is proposed to tackle the poor generalization performance and slow convergence of traditional CNN structures in SAR ATR.In particular,the large margin softmax classifier in the last layer of LM-BN-CNN is utilized to increase the separability of samples.Additionally,batch normalization is utilized to increase the convergence speed.Experimental results have shown that LM-BN-CNN has higher recognition accuracy and better generalization performance than traditional CNN structures designed for SAR ATR.3.To address the problem that redundant feature maps in CNN disturb the classifier and degrade the recognition performance,a novel structure,i.e.the enhanced squeeze and excitation network(ESENet)is proposed.By the attention mechanism,it enhances features containing important information while suppresses redundant features with less information.Experimental results on the MSTAR dataset has shown that the ESENet can automatically extract more efficient features from SAR images and improve the recognition performance.4.To address the problem that the performance of existing deep learning-based SAR ATR algorithms degrades seriously with insufficient training samples,a novel few-shot learning(FSL)framework,i.e.the Conv-Bi LSTM prototypical network,(CBLPN)is proposed,which effectively reduces the amount of training samples for the targets to be recognized.The recognition procedure of CBLPN consists of two main stages.In the first stage,the Conv-Bi LSTM network is trained to map the SAR images into a new feature space where the classification problem becomes easier.In the second stage,a classifier based on the Euclidean distance is utilized to classify samples in the new feature space.Experimental results in the three-class classification problem have shown that CBLPN achieves high recognition accuracy with only a few training samples.5.Most existing deep learning models for SAR ATR utilize classifiers based on inductive inference,which classify the test samples independently during the testing phase.The recognition performance of such classifier primarily relies on the similarity between the training samples and the test samples.In a few-shot SAR ATR task,the training samples are extremely scarce and these inductive inference-based classifiers achieve poor recognition accuracy.To solve this problem,a hybrid inference strategy that combines inductive inference and transductive inference is proposed.Furthermore,a novel few-shot learning framework,namely the hybrid inference network(HIN)is proposed,which overcomes the deficiency of singly adopting inductive inference or transductive inference.The classifier in HIN consists of the inductive inference section and the transductive inference section.In the inductive inference section,each sample is recognized independently according to a metric based on the Euclidean distance.In the transductive inference section,all samples are recognized as a whole according to their manifold structures by label propagation.Finally,in the hybrid inference section,the classification result is obtained by combining the above two inference methods.In order to train the framework more effectively,a novel loss function,i.e.the enhanced hybrid loss,is proposed to constrain samples to gain better inter-class separability in the embedding space.Experiments have shown that HIN effectively improves the accuracy and stability of the few-shot SAR ATR method.