S2F-YOLO: An Optimized Object Detection Technique for Improving Fish Classification

The current emergence of deep learning has enabled state-of-the-art approaches to achieve a major breakthrough in various fields such as object detection. However, the popular object detection algorithms like YOLOv3, YOLOv4 and YOLOv5 are computationally inefficient and need to consume a lot of computing resources. The experimental results on our fish datasets show that YOLOv5x has a great performance at accuracy which the best mean average precision (mAP) can reach 90.07% and YOLOv5s is conspicuous in recognition speed compared to other models. In this paper, a lighter object detection model based on YOLOv5(Referred to as S2F-YOLO) is proposed to overcome these deficiencies. Under the premise of ensuring a small loss of accuracy, the object recognition speed is greatly accelerated. The S2F-YOLO is applied to commercial fish species detection and the other popular algorithms comparison, we obtained incredible results when the mAP is 2.24% lower than that of YOLOv5x, the FPS reaches 216M, which is nearly half faster than YOLOv5s. When compared with other detectors, our algorithm also shows better overall performance, which is more suitable for actual applications.