Automatic marine fish detection using deep learning under laboratory conditions.

Abstract

[eng] In the last decade, deep learning has revolutionized almost every scientific discipline and everyday tasks. In behavioural ecology, deep learning allows us to automatize the acquisition of animal behaviour and improve the analysis of large amounts of behavioural data. Here, we have trained an image-based deep learning algorithm, the Faster R-CNN (Faster region-based convolutional neural network), to automatically detect a marine fish under laboratory conditions, aiming to obtain an automatic tool to study fish behaviour from video recordings. For the training, we have used a total of 14000 labelled images and a data augmentation technique to explore the performance of the fish detection algorithm. Then, we have validated its functioning at different training and augmentation degrees, processing more than 52039 frames for every validation, with and without the presence of the marine fish, Sparus aurata in normal and altered (introduction of a novel object) laboratory conditions. The neural network in its final and best version, trained with all the images and with data augmentation, reached an accuracy of 93%, proving to be a good instrument to study fish behavioural ecology in a non-invasive way.