SAR Object Change Detection Study.

Abstract : This study assessed the applicability of three area-based change-detection algorithms to the problem of recognizing differences between two images. These approaches are based on defining sets of descriptors for small regions of each scene and then comparing these sets of descriptors to indicate areas of change. The first algorithm is based on the phase-correlation concept--a technique for extracting accurate estimates of the relative displacement of two images from the phase of their cross-power spectrum. The 2nd is based on using a subset of power-spectral coefficients for each image--the subset is chosen to minimize differences due to scene bias, shading, or high-frequency noise content. The 3rd is based on defining image moments and a set of functions of these moments which are invariant with respect to a number of geometrical distortion or scale changes of the image. These algorithms were applied to furnished SAR images. Results of the three algorithms are consistent among themselves but correlate weakly with known ground truth. Studies to determine the size of targets that can be reliably detected by use of these methods indicate that the size of targets in the furnished images fails to meet the size criterion for reliable change detection. It is concluded that, for imagery of this quality and resolution, target detection and association algorithms should prove to be more effective for change detection.