Study On Change Detection And Target Classification For SAR Images Under Restricted Information

As an active microwave sensor,synthetic aperture radar(SAR)can form high resolution images with relative invariance to weather and lighting conditions,and has the penetration ability to some extent.Therefore,it has been widely used in military and civilian fields.As the important parts of SAR image interpretation,change detection and target classification have received extensive attention from scholars at home and abroad.However,in practical applications,the available SAR image information is usually limited,which is mainly reflected in the following two aspects: one is that it is difficult to obtain a large number of SAR images,and the other is that the poor interpretability of SAR images makes the target information provided by the images limited and labeling samples difficult.In this dissertation,we carry out the in-depth analysis and research on SAR image change detection and target classification under restricted information.The main contents can be summarized as the following four aspects:1.Aiming at the difficulty in obtaining prior information and sensitivity to the speckle noise in SAR image change detection,we carry out the research and propose an unsupervised SAR image change detection method based on scale invariant feature transform(SIFT)keypoints.The traditional SAR image change detection methods firstly use the SAR images before and after the change to generate the difference image,and then directly analyze the difference image to obtain the change map.In this situation,the speckle noise outside the changed regions will produce large number of false alarms and reduce the detection accuracy.Since the SIFT keypoints are robust to the speckle noise,the proposed method detects the SIFT keypoints from the difference image and exploits the position and scale information of SIFT keypoints to determine the rough range of changed regions,i.e.regions of interest(Ro I).The following detection is around the Ro I,thus reducing the influence of speckle noise outside the changed regions.In addition,different from the traditional methds which processes the blurry difference image,the proposed method make analysis on the original SAR images instead of the difference image to obtain the change map.Therefore,the edge information of changed region can be better preserved.Due to the introduction of key points of SIFT and the utilization of original SAR images,the proposed method can obtain lower false alarms,and at the same time can preserve the integrity of the changed region,thereby improving the detection accuracy.The experimental results based on different measured SAR data demonstrate that the proposed change detection method performs better than the traditional detection method.2.Aiming at the difficulty of labeling samples in SAR target discrimination,we carry out the SAR target discrimination algorithm and propose a semi-supervised SAR target discrimination method based on co-training.Most of the traditional target discrimination methods are fully supervised.That is to say the training samples need to be manually labeled to train the classifier.However,in practical applications,it is usually time consuming and difficult to manually to label the samples.When the amount of samples to be labeled is large,it will take a lot of manpower and time to label them.When only a small number of labeled samples are available,the pattern underlying in the image cannot be captured,resulting in the decrease of discrimination performance.The proposed method can use a small number of labeled samples and a large number of unlabeled samples to implement the SAR target discrimination.Specifically,the existing discrimination features are analyzed,and two sets of features with complementary information are selected as the two views of co-training,based on which two classifiers are trained.During the training stage,one classifier selects its most confident samples to the other to augment the training set.This process is repeated until there are no unlabeled samples.The proposed method utilizes the complementarity between traditional Lincoln features to promote information interaction between classifiers,thus improving the discrimination accuracy.The experimental results based on the measured data set show that using a small number of labeled samples,the proposed method can achieve comparable discrimination performance with the fully supervised method.3.Aiming at the poor discrimination ability of traditional discrimination features and difficulty in labeling samples in SAR target discrimination,we propose a semi-supervised infinite latent dirichlet allocation(SSILDA)model for SAR target discrimination.The traditional discrimination features only provide a rough and partial description for the samples,and lack the description of semantic information.The proposed model can mine the semantic information of the samples and implement target discrimination at the semantic level.Specifically,the proposed model introduces a class variable into the original LDA model,so that the model can implement target discrimination.In addition,the Dirichlet process(DP)is introduced into the model to automatically determine the number of topics.The modified model combines feature learning and discriminator learning into a uniform Bayesian framework,making it possible to extract the semantic features of samples while implementing target discrimination.Thus,the mismatch of features and discriminators can be avoided.Moreover,we introduce the semi-supervised idea into the modified model to reduce the need for the labeled samples.The experimental results based on the measured datasets show that compared with the traditional fully supervised discrimination methods and semi-supervised methods,the proposed method can obtain better discrimination performance.4.Aiming at the limited target information provided by SAR images,we carry out the research on SAR target recognition,and propose a discriminative and generative fusion network for SAR target recognition.The SAR images provide limited information about targerts.On the one hand,different from the optical images,SAR images cannot provide rich information about the structure and color of targets.On the other hand,the SAR images that can be used for training are limited.Therefore,the recognition performance of methods based on deep learning is limited.In practice,the high-resolution range profile(HRRP)can be obtained while obtaining the SAR image.Therefore,we realize the target recognition by simultaneously using SAR images and HRRP data.The structure-related features are extracted from SAR images,and the generation-related features are extracted from HRRP,thus providing a more complete description of targets and improving the recognition performance.Specifically,the SAR image is input into a convolutional neural network(CNN)to extract features related to the target discrimination.The HRRP data corresponding to the SAR image is input into a variational autoencoder(VAE)to extract the features related to the generation of HRRP.Finally,the two features are fused,and the fused features are used for recognition.During the learning process of the entire network,the two tributary networks are jointly optimized.Experimental results based on measured data show that the proposed fusion network has better recognition performance than the existing networks without data augmentation.