Noise resistant estimation techniques for SAR image registration and stereo matching

Presents optimal estimation techniques for extending image registration and stereo matching accuracy under low effective signal-to-noise ratio conditions typical of SAR change detection and stereo applications. Previous algorithms derive multiple local correspondences, which are fit to a transformation model using weighted-least-squares (WLS) methods. The model is refined iteratively by rejecting outliers, i.e. correspondences that produce large residuals in the WLS fit. With the new method, accuracy is improved based on an adaptive optimal estimation formulation. A position error covariance matrix is calculated for each local correspondence (using local image statistics), providing the inverse of the weights for the WLS fit. A vector extension of Huber´s robust regression algorithm utilizes each covariance matrix. In effect, each residual is projected onto the eigenvectors of its covariance matrix and each resulting vector component tested in order to reject (or downweight) outliers. The algorithm is effective even when most of the local correspondences are degenerate. Experimental results show a general improvement in accuracy, in some cases by over an order of magnitude.