Computer vision neural network using YOLOv4 for underwater fish video detection In Roatan, Honduras

Abstract

Fish spawning aggregations (FSA) are groups of individuals of a specie that group for reproduction purposes. Studies show that globally fish species of commercial importance and their reproduction sites are decreasing at an alarming rate. In Honduras there is little information on these reproduction sites, marine biology analysts have determined that it is of crucial importance to monitor these sites.Obtaining video recordings of the FSA’s is vital to study and monitoring efforts, however it turns out to be very time consuming. This study proposes a neural network capable of the detection, classification, and counting of 11 fish species indicatives of reproductive activity at FSA. With this implementation it will be possible to dedicate more time to communication and socialization of the knowledge obtain to local authorities and fishermen, reduce personnel cost, and allow for the inclusion of innovative and technological solutions.Using the Incremental methodology, the Caribbean Fish database of 387 images is created for the training process of the neural network using YOLOv4 and Darknet as infrastructure and architecture respectively with an experimental approach. After 3 stages of training and testing, the object detector and classifier obtained a mAP of 88.1% and average retention of 80%. The network shows that is very capable of complying with the requirements of the monitoring process of fish reproduction sites through video recordings.