Guided block-matching for sonar image registration using unsupervised Kohonen neural networks

This paper proposes an extended block-matching method for registrating side-scan sonar images. Indeed, this work main objective is to exploit block-matching principle and improve its performance by embedding a relevant guidance algorithm. Instead of carrying out the block-matching process on the whole input images, which takes a lot of time, an unsupervised image segmentation step is introduced prior to the matching phase in order to guide it. Thus, the block-matching is only performed on similar regions from the two segmented images, where the potential for finding relevant pairs of blocks is high. This improved version is expected to take less time than the original one. In this work, textural features extracted from both images, feed self-organizing neural networks (Kohonen maps) which implement the unsupervised segmentation step. Experimental results show the effectiveness of the proposed guidance method by reducing registration computation time without any quality loss when compared to the regular block-matching algorithm.