Unsupervised Transfer Clustering For SAR Images Segmentation Based On Dictionary Learning

SAR (Synthetic Aperture Radar) imaging technology has the capability of imaging all-time and all-whether, which makes the interpretation of SAR images become more and more important both in military and civilian usage. And as the first step of image understanding SAR image segmentation plays a more important part in the following interpretation process. Considering the reality that the lablelled SAR image resources are precious and that SAR images have serious speckle noises, the unsupervised transfer learning technology will be used in SAR images segmentation. Meantime data sparse representation dictionary learning method is introduced into our subject. Combined with transfer learning, dictionary learning method can realize knowledge transferring easily and it has great capability of suppressing noises. Three different methods are proposed for our subject. The achievements are as follows.(1) Unsupervised sample transfer clustering method based on dictionary learning to deal with the problem of low separability caused by the low resolution or serious noises. By transferring samples which have great separability from single source images, the proposed method can build a better source classifier to guide and improve the target classifier to get a better segmentation result of target SAR image data. The proposed method is tested on texture images and SAR images segmentation experiments.(2) Unsupervised feature transfer clustering method based on dictionary learning. The key using dictionary method for classification problem is to enhance the discriminative capability of dictionary classifiers. So discriminative atoms will be found in target dictionaries first then helpful feature knowledge will be extracted from multiple source domains with the help of target clustering centers. Next the helpful features will be put into target dictionaries to improve the discriminative ability of target dictionaries. Then we test proposed method on texture images and SAR images.(3) In order to pay enough attention to the uniqueness of individual sample, the dynamic ensemble selection method is introduced. And the data sparse representation method is used to solve the time consuming problem of dynamic ensemble selection method. The UCI datasets are used to test the performance of sparse dynamic ensemble selection. Then transfer sparse dynamic ensemble selection method is proposed based on the sparse dynamic ensemble selection. And the method is tested on SAR images segmentation experiments.This work was supported by the National Basic Research Program of China (No.61003198), and the China Postdoctoral Science Foundation funded project (No.20090460093).