Deep learning based fish length estimation. An application for the Mediterranean aquaculture

Abstract

The aquaculture industry is one of the faster growing sectors of primary production. In recent years, innovative systems and high precision methodologies have been developed, improving all the aspects of production including fish living conditions. Fish biomass estimation is a key factor for husbandry, as it affects operation cost and final production. Here we propose a system that uses a stereo camera and machine vision/artificial intelligence algorithms to estimate the fish size of the Gilthead seabream and the European seabass in aquaculture cages. The system, without human intervention, acquires synchronized images from both cameras and by using a convolutional neural network, detects key measurement points on fish (snout, eye, pelvic fin and fork tail). To train the key-point detection network we created a dataset of manually annotated fish images. Given the fish body-parts (key-points), stereoscopy is used to estimate the fish length. The system was evaluated on a real dataset containing images from cage farms with known fish length distribution. The system achieved high accuracy in length estimation with mean relative length error of Gilthead seabream and the European seabass of 3.15% and 7.4%, respectively.