Research On SAR Image Classification Based On Transfer Learning Algorithm

The classification of the Synthetic Aperture Radar(SAR) images is a kind of image processing techniques that use the feature information of pixels like gray scale, texture, shape, edge and orientation to determine its category in the SAR image and divide different areas accordingly.With the increment of image resolution, there lies massive growth of target information. The conventional supervised classification method for SAR image relies on a large number of labeled observational samples to construct the classification model which costs a lot. In addition, the conventional supervised classification method for SAR image issues that the distribution of the source image and the target image have to remain the same. However, the migration between the scenario of learning and application is more likely in practice. Accordingly, classification errors are likely to occur when the training samples and the testing sample do not belong to the same probability distribution.Consequently, getting precise classification result is one of the main difficulties in the problem of SAR image interpretation.In this paper, the method of transfer learning will be imported in the field of SAR image classification, here gives the main contents and innovations:(1) For the heavy cost of getting abundant labeled observational samples and the troubling migration between the scenario of learning and application, it studied a series of mature transfer learning theories, and the model that combine the theory of active learning and transfer learning. It also studied a series of widely used feature extraction methods and features compression methods for SAR images, and the inquiry mechanisms that could select the most informative samples in the filed of transfer learning.(2) For the inadequate aspect of the traditional transfer learning model in the field of SAR image classification, it studied two types of transfer learning models that suit for SAR image classification. First of all, the principle of transferring risk leaded by the inter-domain differences is discussed. After that the measurement of the similarity between images is studied. At last, two kinds of existing transfer learning methods for remote sensing images is studied and applicated to SAR image classification.(3) For the insufficient storage and computation problems for transfer learning algorithm based on traditional SVM classifier, it proposed a sample reusable transfer learning algorithm based on SVM classifier relies on its own characteristics. It imported a sample recycling mechanism in the target domain and a sample dismissing mechanism for invalid sample in the source domain. The experimental results show that this method improved the utilization efficiency for training samples and optimized the classification results for SAR images.