MT-YOLO: Underwater Fish Detection Model Based on YOLOv8s

Underwater fish detection is important for underwater ecological conservation and underwater species monitoring. In this study, an underwater fish recognition method, MT-YOLO, is proposed based on the YOLOv8s model and trained and tested on a self-constructed underwater fish dataset containing 12,000 fish targets. Firstly, in order to enhance the feature fusion capability of the original network, an underwater multi-scale target fusion module (MF-Module) is designed in this study, which improves the model's ability to process underwater multi-scale targets by receiving input and output features from different scales; second, a multi-tasking module (TP-Module) is designed in this study, which improves the multi-tasking capability of the original network by using multiple detection header to improve the multitasking performance of the original network, and the module also introduces shared convolution to reduce the number of parameters of the model. Experiments show that MT-YOLO increases the accuracy by 4.8% and reduces the model size by 15.6%. Meanwhile, this study ablates MT-YOLO with underwater detection algorithms such as SDD The experimental results show that MT-YOLO has stronger performance in detecting fish in complex scenes.