Image segmentation is a fundamental issue in image processing. Segmentation of synthetic aperture radar (SAR) images is extremely difficult on account of intrinsic multiplicative speckle noises. Due to the ambiguities of SAR images, labeled instances are difficult and time-consuming to obtain while unlabeled data are abundant. In this paper, a new semi-supervised approach based on transductive support vector machine (TSVM) is proposed to segment SAR images, it is robust to noises and is effective when dealing with low numbers of high-dimensional samples, moreover, it could efficiently make use of unlabeled data to reduce human labor and improve precision. Segmentation results are also compared to SVM and TSVM trained by using different samples and parameters. Experimental results demonstrate that the proposed method is very promising.
- SAR;image segmentation;semi-supervised learning