Underwater fish detection and counting using image segmentation

There is a great demand for precision fish farming in aquaculture, and accurate fish counting is essential for precise control and monitoring of fish farming processes. Traditional fish counting is manual, time-consuming, and labor-intensive. Image recognition is an effective method for automatic counting. However, the accuracy of automatic detection in real environments is limited by the different sizes, swimming styles of fish, and the potential for mutual occlusion between multiple moving objects. This study introduces a BoTS-YOLOv5s-seg model based on YOLOv5s to achieve accurate detection of fish edge by using the YOLOv5s instance segmentation model and improving the loss function to SIoU loss while improving non-maximum suppression to reduce missed detection of overlapping objects, and finally incorporating a bottleneck transformer to make the model more focused on valid image information and reduce the model parameters. The BoTS-YOLOv5s-seg model proposed in this study to a realistic scenario with a farmed fish dataset had better performance and resulted in a small model size of only 7.1 M, and the values of mAP@0.5 and GFLOPs of the proposed algorithm reached 90.9% and 25.4, respectively, surpassing the traditional YOLOv5s-seg, YOLOv5m-seg, and YOLOv5l-seg in terms of detection results. This research proposes a method that can effectively support accurate fish counting in the context of intelligent aquaculture.