An Improved YOLOv8 Fish Identification Based on PConv and Attention Mechanism

Abstract

Accurate and effective identification of fish species is crucial for the processing and transportation of fish products. It enables efficient tracking and management of fish products, making it an essential aspect of the industry. Traditional methods of fish identification usually require expert knowledge within a particular domain. However, these methods can struggle to capture complex fish characteristics, particularly under varying lighting and angle conditions. In the article, an effective recognition algorithm named YOLO-FD is proposed. First, a novel feature extraction module has replaced the C2f module in YOLOv8, which effectively reduces the amount of model parameter calculations. The Efficient Multi-Scale Attention (EMA) applied in this study adeptly preserves spatial and channel information, fostering inter-regional interaction and enhancing feature extraction within the backbone network. The loss function in YOLOv8 was improved to address the sample imbalance problem. The YOLO-FD as an effective fish recognition algorithm addresses challenges faced by traditional methods simultaneously enhancing identification accuracy and offering lightweight improvement.