Underwater Image Registration

Abstract

Underwater image registration is widely needed for various applications these days. For example, to access the minerals and other resources present in the oceans and seas, to explore and examine the biodiversity in marine environment, to supervise the dam structure, oil refineries and cables passing through sea, to study seabed, etc. Registration of underwater images is a challenging task as the images captured underwater are degraded. The absorption of light in water, scattering of light due to particles present, addition of noise are some of the causes of degradation of underwater images.
Research has been carried out widely in the area of image registration. For the feature detection and matching, Scale Invariant Feature Transform (SIFT) is the technique used most of the times. SIFT is an algorithm for feature detection and matching. The features extracted in SIFT are invariant to scale, rotation change in illumination and addition of noise. Due to poor quality of underwater images, SIFT algorithm extracts less number of features hence reduces the accuracy of image registration.
In order to achieve better registration of underwater images, improvements in SIFT algorithm is proposed. As the underwater images gets affected by noise, filtering of image is proposed before applying SIFT algorithm to it. Gabor filter is used as prefilter. SIFT algorithm eliminates the key points with low illumination as there is a large possibility of it being selected from the background. To eliminate these points, pixel value 0.03 is used as threshold. But as most of the underwater images have less illumination, this rejects many key points. Hence it is proposed to make the threshold value for low illumination point elimination adaptive. 10 percent of the total contrast of an image is considered as the threshold value.
Descriptor of key point plays a vital role in matching of key points. For a descriptor to carry more and correct information, it is proposed to apply sobel operator before forming a descriptor. To obtain match between the key points, SIFT algorithm calculates the Euclidean distance. But is it observed that it gives some false matches. Hausdorff distance is proposed to use for calculation of distance between two key points. More number of key points is detected and proper matching is achieved with this modified algorithm. This gives a better registration of underwater images.