SAR Image Target Recognition Based On Sparse Representation And Dictionary Learning

Synthetic Aperture Radar(SAR)is an active sensor,which can work all-day and allweather,so it is widely used in many fields such as disaster assessment and military defense.With the use of high-resolution synthetic aperture radar,large-scale images with complex features are generated,manual interpretation is difficult,so that Automatic Target Recognition(ATR)of SAR image becomes a research hotspot.In this paper,SAR image target recognition technology based on sparse representation and dictionary learning is studied according to SAR image characteristics.The main research work of this paper is as follows:(1)A SAR image preprocessing method is proposed.Firstly,the logarithmic transformation of images is performed to suppress the noise.Then the two dimensional maximum entropy threshold segmentation method is used to segment the target region and the shadow region.A method based on the genetic algorithm for calculating the optimal segmentation threshold is proposed.Finally,the target region is processed,including image enhancement,centroid registration and grayscale normalization.The experimental result shows that the preprocessing process is effective and the classification information of SAR image is enhanced.(2)A SAR image target recognition method,NN-ESR(Nearest Neighbor and Extended Sparse Representation),is proposed,which combines nearest neighbor and extended sparse representation.Firstly,SAR images are preprocessed.Secondly,two dimension principal component analysis(2DPCA)is used to extract the feature vectors of the image,and a method for selecting feature vectors according to recognition ability is proposed.Finally,the classification of the SAR image target is determined.In order to solve the noise and occlusion problem of images,the traditional sparse representation model is extended.What’s more,in order to improve the recognition speed of the extended sparse representation,NN-ESR combines nearest neighbor method and extended sparse representation.(3)A SAR image target recognition method based on dictionary learning and extended joint dynamic sparse representation,DL-EJDSR(Dictionary Learning and Extended Joint Dynamic Sparse Representation),is proposed.First of all,in the step of image preprocessing,the target area and the shadow area are segmented.Combining the information of these two areas can represent the image better.Secondly,instead of directly using all the training samples as fixed dictionaries,a dictionary learning method,LC-KSVD,is introduced into the training phase to learn the feature dictionaries of target area and shadow area.Finally,the extended joint dynamic sparse representation algorithm,EJDSR,is proposed in the testing phase.It is used to solve the sparse representation coefficient,and the target category is determined according to the reconstruction error.In order to validate the SAR image target recognition methods proposed above,comparative experiments are carried out on the standard data set.The experimental results show that,compared with the sparse representation classification method,although the recognition rate of NN-ESR is slightly reduced,but the recognition speed is greatly improved.And DLEJDSR makes different classes more distinguishable and effectively improves the target recognition accuracy.Therefore,NN-ESR is suitable for online image recognition,while DLEJDSR is suitable for offline image recognition.