Comparison of YOLOv7, YOLOv8, and YOLOv9 for underwater coral reef fish detection

Abstract

Automated underwater fish detection offers an appealing solution to improve efficiency and cost-effectiveness compared to labor-intensive manual detection methods. This study conducted a thorough assessment of three state-of-the-art single-stage detectors belonging to the You Only Look Once (YOLO) series – namely, YOLOv7, YOLOv8, and YOLOv9 – focusing on the detection and classification of four dominant coral reef fish species. These YOLO models were trained using a customized dataset comprised of underwater images showcasing the fish species, sourced from Pulau Bidong and neighboring islands in Terengganu, Malaysia. Data collection was facilitated using the Stereo-Diver Operated Underwater Video System (Stereo-DOVs). The main objective of this study is to determine the top-performing model for precisely detecting and classifying the fish in the images. Notably, each of the YOLO models achieved high mean Average Precision (mAP)@0.5 scores, with percentages of 96.6%, 97.9%, and 94.3% respectively. Further visual examination showcased the models’ adeptness in accurately detecting the majority of fish instances within the test dataset and dataset images from the internet, confirming their robust performance. Taking into account both the evaluation metrics and visual results, YOLOv7 and YOLOv8 stand out as appealing choices to be used as the base models for our future study.