Segmentation of SAR images by means of Gabor filters working at different spatial resolution

The traditional approaches to contour detection in SAR images employ a priori models of the speckle. The authors propose to apply to SAR images a new contour detection algorithm which is intended to be data-independent, i.e., it does not exploit any domain specific knowledge about the contents of the scene. This new contour detector must feature some degree of biological plausibility in order to simulate the behavior of a human photo-interpreter who is able to segment at one glance images acquired by optical as well as by SAR sensors. This means that, in order to improve the performance of the contour detector, the authors are not exclusively restricted to the exploitation of SAR images, but they can apply the algorithm to a wide variety of test images which refer to well known physiological and psychophysical experiences. After the contour detection stage, they employ a second stage which extracts (closed) regions from non-connected binary contours, and a third stage which performs (conservative) region-merging. The third stage employs a neural network clustering algorithm whose input sequence consists of segment-based textural parameters. The statistical regularities detected by the network in the random input sequence are equivalent to categories of segments which can be successively used by a knowledge-based, domain dependent, hierarchical classification scheme. The proposed segmentation procedure is applied to SAR images depicting sea ice. The performance of the contour detection stage seems to encourage the development of a new category of biologically plausible algorithms to perform low-level visual tasks