Fish species recognition using computer vision and a neural network

Abstract

A system is described to recognize fish species by computer vision and a neural network program. The vision system measures a number of features of fish as seen by a camera perpendicular to a conveyor belt. The features used here are the widths and heights at various locations along the fish. First the measured values are used as input values to a neural network, together with the information on the species. The network is trained to recognize the species from these input data. To decrease the time to train the network, a learning rate, a momentum factor and the elimination of non-contributing connections and nodes were introduced. Testing of the network showed that more than 95% of the fish could be classified correctly.