Wavelet based similarity measurement algorithm for seafloor morphology

Abstract

The recent expansion of systematic seafloor exploration programs such as geophysical research, seafloor mapping, search and survey, resource assessment and other scientific, commercial and military applications has created a need for rapid and robust methods of processing seafloor imagery. Given the existence of a large library of seafloor images, a fast automated image classifier algorithm is needed to determine changes in seabed morphology over time. The focus of this work is the development of a robust Similarity Measurement (SM) algorithm to address the above problem. Our work uses a side-scan sonar image library for experimentation and testing. Variations of an underwater vehicle's height above the sea floor and of its pitch and roll angles cause distortion in the data obtained, such that transformations to align the data should include rotation, translation, anisotropic scaling and skew. In order to deal with these problems, we propose to use the Wavelet transform for similarity detection. Wavelets have been widely used during the last three decades in image processing. Since the Wavelet transform allows a multi-resolution decomposition, it is easier to identify the similarities between two images by examining the energy distribution at each decomposition level.

(cont.) The energy distribution in the frequency domain at the output of the high pass and low pass filter banks identifies the texture discrimination. Our approach uses a statistical framework, involving fitting the Wavelet coefficients into a generalized Gaussian density distribution. The next step involves use of the Kullback-Leibner entropy metric to measure the distance between Wavelet coefficient distributions. To select the top N most likely matching images, the database images are ranked based on the minimum Kullback-Leibner distance. The statistical approach is effective in eliminating rotation, mis-registration and skew problems by working in the Wavelet domain. It's recommended that further work focuses on choosing the best Wavelet packet to increase the robustness of the algorithm developed in this thesis.