Research On Sar Images Classification Based On Active Learning

SAR image classification is an essential part and key technology for SAR image interpretation, but the existing methods rely too much on the amount and the quality of the available reference samples. SAR images do not look like what they are so the labeling procedure costs very much. So how to exploit more useful information from a limited number of labeled samples to reduce labeling cost urgently needs to be solved. In this paper, we focus on the efficiency of Active Learning when applied to the field of SAR image classification. The main work accomplished in this paper are as follows:1) We study the standard and procedures of the labeling process. The labeling of samples is one of the most delicate phase for the success of SAR image classification though it is usually ignored by researchers. In this paper, we summarized the standard and procedures of the labeling process and detailed the steps to solve the ambiguity problems we may encounter.2) We propose a novel AL SAR image classification based on local area after taking a full consideration of the cost difference between labeling loads of different samples. By quantifying the labeling cost of different samples, we can select the informative samples using this constrain. So the labeling cost can be reduced further.3) By incorporating a constrained semi-supervised learning method, we propose a novel transfer learning method. Thus, we can make maximum use of the unlabeled samples in the target domain to avoid human labeling every iteration. And the result is the reduction of human labeling load.4) Implement a SAR image classification software platform. Based on the AL and transfer learning we propose in this article, we implement a software which can do some basic image processing. We take a full consideration of scalability of the platform.