Identifying coral genera using median binary partition texture patterns

In this paper, underwater still images taken using the transect line method by marine scientists, were used to distinguish live coral textures from rubble/dead coral textures and sand textures. The live coral textures were further classified as either, Acropora, Montipora, Favia, Goniopora, or Porites, the top five coral genera found in the Philippines. A collection of 250 training texture images were used-50 for each coral genus, 131 samples for other coral types, 100 for sand, and 100 for rubble. Image processing (color normalization and median binary partition (MBP) for texture description) were conducted on the images. K-nearest neighbor (k-NN) classification was then used to classify the different texture regions. MBP proved useful in describing pixel relationships among texture regions, and when combined with k-NN, required less computational resources in classifying coral textures. Coral detection had a rate of 92.6 percent while sand and rubble/dead corals texture had detection rates of 93 percent and 84 percent, respectively. The effectiveness of MBP was reduced due to scale and multi-resolution limitations of the set of acquired images used for training and testing among the live corals classified, Porites had the highest detection rate (86%) followed by Acropora (58%), Montipora (54%), Favia (52%), and Goniopora (42%).