ENHANCING UNDERWATER LIVE FISH DETECTION PERFORMANCE WITH YOLOV11 AND IMAGE ENHANCEMENT TECHNIQUES IN A REAL-TIME SYSTEM

Abstract

Underwater fish detection is challenged by low light, turbidity, and blue-green color dominance from light attenuation. This study aims to compare six image-enhancement scenarios (baseline, CLAHE, Retinex Ultra Lite, UDP, UDP Super Lite, and Gamma Correction + White Balance) combined with YOLOv11 to evaluate their detection accuracy and inference speed under challenging underwater conditions, and identify the most suitable method for real-time embedded underwater vision. Experiments used a controlled fish tank, filmed with an OAK-D camera under five natural lighting conditions (09:00–16:00 WIB). Performance was assessed qualitatively and quantitatively on a 600-frame test set (four illumination conditions), computing precision, recall, and mAP@0.5 from YOLOv11 predictions, plus inference speed in frames per second (FPS) on a Raspberry Pi 5 + OAK-D. CLAHE gave the best balance (precision 0.86, recall 0.83, mAP@0.5 0.84, 17 FPS); Gamma + White Balance was slightly less accurate (mAP@0.5 0.82) but faster (19.8 FPS); both outperformed the baseline (mAP@0.5 0.69, 21.3 FPS). Retinex Ultra Lite, UDP Super Lite, and full UDP all fell below 15 FPS (5.8–8.7 FPS), making them unsuitable for real-time edge deployment despite similar or lower accuracy. Lightweight, effective enhancement significantly improves real-time underwater fish detection on embedded devices, with CLAHE identified as the most suitable method for underwater object identification and quantification systems.