Visual seabed classification using k-means clustering, CIELAB colors and Gabor-filters

Abstract

In this article, we discuss visual classification using unsupervised learning combined with methods that originate from human vision to divide the Baltic seabed to the soft and hard areas.

Seabed classification plays an important role in an understanding the undersea environment. Seabed can be characterized to be as muddy, rocky or sandy. Mine countermeasures (MCM) missions normally are clearance and/or route finding types and in both of these cases successful detection and classification is strongly connected of seabed type.

As our unsupervised learning method, we used k-means clustering. When we filtered our gray-scale seabed picture using Gabor filters, we noticed significant improvement after we segmented filtered image with k-means. We will also show results that we achieved using k-means alone and with Lab colors that are designed to approximate human vision.