Fish classification using extraction of appropriate feature set

Abstract

The field of wild fish classification faces many challenges such as the amount of training data, pose variation and uncontrolled environmental settings. This research work introduces a hybrid genetic algorithm (GA) that integrates the simulated annealing (SA) algorithm with a back-propagation algorithm (GSB classifier) to make the classification process. The algorithm is based on determining the suitable set of extracted features using color signature and color texture features as well as shape features. Four main classes of fish images have been classified, namely, food, garden, poison, and predatory. The proposed GSB classifier has been tested using 24 fish families with different species in each. Compared to the back-propagation (BP) algorithm, the proposed classifier has achieved a rate of 87.7% while the elder rate is 82.9%.