Prediction fat percentage and visceral weight from whole fish images with a multi-input neural network

In aquaculture, high accuracy in trait measurements benefits the genetic progress from a breeding program. Breeding traits like fat percentage and visceral weight are related to feed/cost efficiency of growth and product quality, and important metabolism and health indicators. Problems concentrate on finding the proper methods to accurately measure or predict these traits, as most current approaches are invasive, labour-intensive or may disturb or damage the fish. Interior trait prediction from image analysis would allow a real-time, large-scale and non-invasive alternative for such traits. This study investigates using whole-fish images in combination with exterior traits to improve the prediction of fillet fat percentage and visceral weight. The result of including images as extra input shows improvement on the accuracy of fat percentage prediction. The neural network extracted contour-based features and brings into view several biological indicators that appear to be informative for prediction.