Segmentation Methods for Synthetic Aperture Radar

Abstract

Synthetic aperture radar (SAR) image segmentation is a fundamental problem in SAR image interpretation. SAR images often contain non-texture object and texture object. Level set method, known as deformable model, is a powerful image segmentation technique. It can get accurate contours of non-texture object, but has poor performance in getting contours of texture object. In recent years, parametric active contour model (snake), which was proposed by Kass in 1987 [1], has become one of the most studied techniques for image segmentation. The snake approach is based on deforming an initial contour or surface towards the boundary of the object to be detected. The deformation is obtained by minimizing a global energy designed such that its minimum is obtained at the boundary of the object. The energy is basically composed by a term which controls the smoothness of the deforming curve and another one which attracts it to the boundary. However, the classical active contour model presents several limitations. In particular, it is sensitive to initial contour placement, and most importantly, it can’t handle topological changes of the curves during their evolution [2-4]. Geometric active contour model was subsequently proposed by Osher and Sethian in 1988 [5]. This model is based on the theory of curve evolution and geometric flows, which implemented using level set. Level set is designed to handle problems in which the evolving interfaces can develop sharp corners, change topology and become very complex. Level set approach has been widely applied to image processing [6-9]. SAR image segmentation is an important, challenging problem and a necessary first step in image analysis and interpretation. However, segmentation of distinct areas, such as city and river, is a challenging task due to their complex topologies. So we use level set approach to solve the topology problem. One class of image segmentation is object detection, where certain objects in the image are to be singled out. In this case, the image is basically divided into two sets: objects and background. Some non-texture objects in SAR image, such as river, ravine and railway, can be detected easily by level set method because of their distinctness with background and lack of texture. But in particular, objects in SAR image often contain texture, such as the city zone, which may cause considerable difficulties when applying level set approach. It makes the detection result too minute, and loses the consistency of objects. We would like to use level set method to get accurate contour of objects, whereas the consistency of objects is also needed. As a result, a new technique which could make level set method adaptable to detection of texture object is required.