RESULTS: Image Registration
Construction of ocean floor photo-mosaics is a challenging task, since it consists in stitching several thousands of sea bottom images into the single consistent mosaic. The core routine of the mosaic building process is the automatic matching of two partially overlapping images. Automatic matching of any kind of images usually presents some difficulties. Underwater images pose even more problems, since they normally suffer from the acquisition noise, object occlusion, parallax effects, non-uniform illumination, lack of image texture and presence of moving objects. For a long time underwater mosaicing has being traditionally performed on video sequences using correlation-based matching methods. However, when transformation between images is much more complex than a simple translation or a big lighting variation is present, the correlation-based methods fail. When used in combination with other techniques, providing larger invariance, the computational load of correlation-involving algorithms increases greatly.

Image Pair

During the last five years several efficient feature-based matching techniques were developed, but mainly for overland applications. For example, SIFT (scale invariant feature transform) algorithm, presented in 2004 by Lowe, SURF (speeded-up robust features, Bay, 2006) and other similar methods demonstrate high level of invariance to image scaling, rotation, change in illumination and 3D camera viewpoint, being at the same time fully automatic and quite fast. The new algorithms were thoroughly tested on various images of human-made environment and natural overland scenes, where they proved to be repeatable, distinctive and robust. However, application of these techniques to more difficult underwater images occured to be not straightforward.

Image Registration The goal and essence of our work was to test recent techniques on various underwater image sequences, to find out the core parameters influencing they performance and to adapt techniques to underwater photometric and geometric artifacts. As a result a specialized for underwater imagery registration process was developed. The process includes preprocessing of images, detection of salient points in images, description of the detected points using the surrounding them image patches, matching of the descriptors, iterative estimation of motion between images and final registration of two images into the single coordinate frame.

The developed registration algorithm introduced significant improvements to the photo-mosaics construction process and enabled registration of challenging underwater images, suffering substantially from blur, loss of color, low contrast, non-uniform illumination and other common for water medium photometric artifacts.