Analysis of feature matching performance on correlated speckle image pairs

The feasibility of feature-based co-registration of synthetic aperture sonar images and other speckled imagery relies on finding a sufficient number of feature matches with sub-pixel localization accuracy. This paper considers the effect of speckle decorrelation on feature matching performance, quantifying these trends by means of simulating correlated pairs of speckled images. Firstly, using a perfectly bland amplitude scene, the number of inlier feature matches at a given speckle coherence was observed to be Poisson distributed. The expected number of inlier matches for a range of coherence values was estimated, demonstrating a rapid decline in matched features as coherence drops. The concept of “feature repeatability under correlated speckle” is introduced as the expected ratio of inlier features to raw feature matches with respect to coherence. The feature repeatability was also estimated when using sand ripple scenes instead of a bland scene. A two-parameter approximate model of feature repeatability is proposed, showing a good fit for both the bland and non-bland cases. Non-bland scenes were found to have significantly more raw features, as well as a higher feature repeatability at any level of coherence, with the repeatability trend for a bland scene acting as a baseline that can be used for prediction. Approximate distributions of the localization error of feature matches were also computed, where the errors were marginally lower for non-bland scenes.