Real-time underwater sorting of edible fish species

Common carp (Cyprinus carpio), St. Peter's fish (Oreochromis sp.) and grey mullet (Mugil cephalus), were sorted according to species while swimming in pond water containing algae and suspended sediments. Fish images were acquired by a computer vision system while swimming through a narrow channel with their sides to the camera so that distance from the camera was relatively constant. Background illumination was used to overcome water opaqueness and to generate high image contrast. An algorithm extracted size- and orientation-invariant features from the fish silhouettes. Classification of the grey mullet, St. Peter's fish and carp images was achieved with a Bayes classifier, to accuracies of 98.9%, 94.2% and 97.7%, respectively. A real-time underwater computer vision system was tested in a pool in which fish swim through a narrow transparent unidirectional channel. Two sets, of 1701 and 2164 images, respectively, were analyzed with overall species recognition accuracy of 97.8% and 98.9%.