Anomalous white shrimp detection in intensive farming based on improved YOLOv8

Abstract

Timely detection of anomalous shrimp is crucial for ensuring farming safety. Therefore, this study developed an effective model to detect abnormal shrimp behaviors, including curling, floating, leaping, cannibalism, and death. The proposed model uses YOLOv8 as the baseline, adjusts network parameters to align with the characteristics of abnormal shrimp, employs content-aware reassembly of features (CARAFE) to preserve more semantic information, and utilizes dynamic convolution to enhance the network's expressiveness. Achieving a 97.8 % mAP@0.5 and 96.1 % F1 score on a custom dataset, the model demonstrated superior detection performance and a smaller size compared with Faster-RCNN, single-shot multi-box detector (SSD), YOLOv5, YOLOv6, and YOLOv7. Based on the proposed model, we developed an abnormal shrimp monitoring system with significant potential to benefit white shrimp cultivators.