AUTOMATIC FISH POPULATION COUNTING BY ARTIFICIAL NEURAL-NETWORK

A new method of automatically counting fish using an artificial neural network is presented. A back propagation of error feed-forward neural network has been trained to count synthetic fish populations. Trained networks are subsequently shown to generalise well to previously unse en fish tank scenes, giving a 94% success rate on scenes containing up to 100 fish in a variety of orientations and overlaps. This out-perfo rms both pixel counting and energy estimation methods.