A Real-Time Hilsa Fish Identification System Based on YOLO Deep Learning Technique

Abstract

Hilsa is one of the commercially important fish species, Tenualosa ilisha and T. toli (Hamilton, 1822) as well as Hilsa kelee need an accurate identification for the sustainable management of marine fisheries. Classical identification of species through morphology is labour intensive and subject to human error, especially for closely related species. In this study, We developed a fish recognition system in real time with use of YOLO (You Only Look Once) deep learning method built to believe Tenualosa ilisha, T. toli, Hilsa kelee and duplicate hilsa based on the main morphological characters. The approach adopted was YOLOv8 model training based on 2,850 annotated images and being trained for a split ratio of 70:20:10 divided among the train-validation-test dataset samples. Rotation and horizontal flipping, box reshaping, brightness adjustment are augmented to make the model more robust. The system appropriately recognized species-specific morphological traits such as body depth ratios, dorsal fin position, structure of the operculum and patterning pattern. Results: Overall identification accuracy was 94.3% (phenotype-averaged mean AP at threshold IoU = 0.5: 91.8%). Identification at the species level for Tenualosa ilisha (96.2% accuracy), T toli, (92.7%), Hilsa kelee (91.4%) and overboard hilsha (93.8%). The developed system offers operational tools for the automated species identification to be used in fisheries surveillance, market control and biodiversity conservation in Indo-Pacihic hilsa habitats.