Unsupervised Texture-Based SAR Image Segmentation Using Spectral Regression and Gabor Filter Bank

Segmentation of synthetic aperture radar (SAR) image is a difficult task in remote sensing applications due to the influence of the speckle noise. Most existing clustering algorithms suffer from long run times. A novel unsupervised segmentation algorithm has been proposed in this paper, based on Gabor filter bank and unsupervised spectral regression (USR), for SAR image segmentation. In the proposed algorithm, we use a Gabor filter bank to decompose the image to several sub-images. Features are extracted from these sub-images and further, learned, using USR. Finally k-means clustering is employed and the image is segmented. The segmentation results were tested on simulated and real SAR images, texture images, and natural scenes. The results of segmentation on texture images show that proposed algorithm has the ability to effectively manage large-size segmentation cases, since the eigen-decomposition of the dense matrices is not required in USR. Meanwhile, the proposed algorithm was more accurate than all of the other compared methods. The running time in MATLAB was compared against parallel sparse spectral clustering (PSSC) and although our proposed algorithm is serial, it had significantly shorter run time compared to PSCC. It is also demonstrated that the clustering of features improves significantly after learning.