SAR Image Segmentation Based On Support Vector Machine

SAR, as an alternative means of long-distance-to-earth observation, with high-resolution, all-weather, the advantages of transmission, in remote sensing, military, hydrology, mining and other fields, have been rapidly developed, and therefore the analysis of the SAR image processing is becoming increasingly important, in which SAR image segmentation is an important aspect of this. The current segmentation methods in the treatment of high-dimensional features, small sample areas are difficult to obtain good segmentation results, mainly for the poor performance of the promotion of capacity, or learning slowly, difficult convergence. In addition, SAR is a coherent imaging system, and it is easilyaffected by speckle noise. In this paper, support vector machine is used in SAR image segmentation. We divide the original image into a small piece of N×N image (a regional), then calculate the three local statistical texture features and the eleven gray scale image block co-occurrence matrix features of each sub-regional image. All the values obtained are scaled into [-1,1], then are sent to SVM to be trained as the prescribed format as training input sample to get the final training good classifier on the SAR image segmentation. An improved one-against-one approach is used to extend two-class SVM to a multi-class SVM, thus multi-target SAR image segmentation is achieved. In the single, multi-target SAR image segmentation process, we use different parameters and different feature combinations to find the best classifier so that we can get better segmentation results and lower error rate. Also a comparison of different segmentation approach is used to show the different effect of a SAR image. The analysis shows that this method is better than many existing methods both in the accuracy and the anti-noise.