Intensity Normalization of Sidescan Sonar Imagery

Sonar imaging is currently the exemplary choice used in underwater imaging. This is because sound waves penetrate larger distances than electromagnetic waves underwater. However, since sound signals are highly absorbed by water, an image acquired by a sonar will have gradient illumination change depending on the distance each wave has travelled. In this work, we investigated this phenomenon, which we refer to as echo decay, and proposed methods to normalize the images with regards to the received acoustic signals (also referred to as illumination in digital image processing context). Two sidescan sonars have been used to capture the seabed in in Lake Vättern in Sweden in two different directions, west-east and east-west, thus a difference of 180º between overlapping regions. This imaging direction difference makes the task extremely challenging due to differences in the acoustic shadows. To verify the appropriateness of the proposed methods and compare them to the non-normalized images, we performed similarity analyses between (several) corresponding regions extracted from each overlapping pair of sonar images. These pairs have rotation differences and slight scaling differences (the scaling was obvious while manually extracting the overlapped regions), not to mention the possible error due to the manual extraction of each pair from the overlapping images which was done by the first and the third authors of this work. Hence, two sets of pairs have been extracted one before illumination normalization and another after normalization. The normalization was performed using MIxed exponential Regression Analysis (MIRA) estimated from each image that needs normalization (MIRA is proposed in this work)). Another normalization method was also implemented using an adaptation of a well-known fog removal technique to sonar imaging, known as Dark Channel Prior Method (DCP). Experiments using the nonnormalized images (i.e. originals), normalized images with MIRA, and the normalized images with DCP have been performed using SSIM similarity score, and a novel global quality metric specifically developed for side-scan sonar images (SIQE) showed that MIRA has superior normalization performance. We ought to mention that this work has been carried out as part of SWARMs project (http://www.swarms.eu/).