An important application in sonar research is change detection, which is reliant on accurate alignment of repeat-pass images. Although less accurate than correlation-based methods, feature-based methods are gaining popularity in radar and sonar imaging due to their relative computational efficiency. This thesis explores the feasibility of image registration of speckled imagery using feature matching.

As a proof of concept, a feature-based registration method is proposed for synthetic aperture sonar (SAS) based on an ideal sonar track. The registration pipeline uses the Scale Invariant Feature Transform (SIFT) and RANSAC estimation, a popular combination in computer vision. Using a novel track registration approach, feature correspondences are used to estimate a set of track registration parameters from which an image registration can be computed. This method produced an accurate alignment to within 0.03 pixels for a simulated repeat-pass scene with a 0.2m baseline.

Desirable aspects of feature matching performance include a sufficient density of detected features, a high ratio of inlier matches, and accurate feature localisation. Since the performance of feature matching is known to be situational, simulated data was used to produce a large number of images from which general trends could be observed. The effect of sinc-interpolated sub-pixel shifts on feature matching performance was measured, with non-oversampled speckle images yielding an inlier ratio of less than 1% in the worst case. Features were signi cantly more robust for oversampled images, with the expected worst-case inlier ratio being around 45% for two-times oversampled images and around 77% for four-times oversampled images. Overall, oversampled images were shown to provide better feature repeatability, increased density of features, and lower correspondence localisation errors compared to images without oversampling.

SIFT and SURF (Speeded Up Robust Features) were evaluated for performance in relation to speckle decorrelation and scene content. Feature matching was shown to be problematic for bland scenes with coherence below 0.9. Feature matching on non-bland scenes yielded more features, increased feature repeatability, and improved localisation accuracy. A model is proposed to capture the trend between feature repeatability and scene coherence for any given feature matching pipeline. Although the feasibility of feature matching in a given scenario cannot easily be predicted, a low number of matches and moderate localisation errors are both likely to have unfavourable implications on success and/or registration accuracy.