Research On Methods Of Target Region Extraction For High-Resolution SAR Images

Using the knowledge of image processing,machine learning and pattern recognition,to investigate the techniques of acquiring target regions from Synthetic Aperture Radar(SAR)images accurately has caught researchers' wide attention.According to the SAR automatic target recognition(ATR)system,the extraction of target region consists of detection and discrimination.This thesis focuses on the studies of the SAR target detection and target discrimination by introducing techniques of the object proposal generation,the mid-level features representation and weak supervised learning(WSL).The main works of this thesis are as follows:1.In chapter 3,a modified Constant False Alarm Rate(CFAR)algorithm based on object proposals for the ship target detection problem of multiple ship target with different sizes in SAR images is introduced.This method firstly learns the gradient-based object proposal generation mode based on the training data with object annotation,then generates a small set of object proposals with different sizes,and finally uses the proposal-based CFAR detector,where the extracted object proposals are regarded as the guard windows instead of setting fixed guard window,to detect the true positive object proposals.By introducing the object proposals,it could gain good detection performance in the multiscale situation,and also could directly obtain the accurate target regions to avoid the problem caused by the target clustering on the pixel-level detection results.The effectiveness of the proposed method is verified using measured SAR data by comparing with the traditional SAR ship target detection methods.2.In chapter 4,a target discrimination method for high-resolution SAR in complex scenes is described by combining the advantages of the mid-level feature generation and WSL.The weak label is indicating whether a detected image contains the target of interset or not,i.e.,the image-level labels.Firstly,the two-parameter CFAR and target clustering method are used to obtain the candidate regions.The all candidate regions extracted from the negative training images are clutter false alarms(CFAs).The candidate regions extracted from the positive training images contain not target regions but also CFAs.Then,the mid-level features are used to represent each candidate region.Based on the generated mid-level features,the WSL approach uses the unsupervised latent Dirichlet allocation(LDA)to initialize a set of target training examples extracted from positive images and starts iterative learning of the target discriminator.Finally,the final target discriminator is used to discriminate the candidate region obtained from the test images.By the introduction of WSL,the target discriminator can be only obtained via the weak labeled training data,which can reduce the manual work greatly.In addition,comparing with the traditional SAR target discrimination features,the mid-level features can better discriminate the targets and CFAs,for providing a better description of the spatial and structural information of the image.The comprehensive and specific experiments on the measured SAR data have demonstrated the effectiveness of the method.