An efficient and lightweight detection method for floating residual feed particles in aquaculture

The detection of floating residual feed particles through computer vision is crucial for achieving smart feeding in aquaculture. Nonetheless, due to the diminutive size and aggregation of the feed particles, as well as the variety of breeding environments, feed particle detection faces difficulties such as low detection accuracy, high computational consumption, and large model size. To address these problems, a novel model, inspired by the You Only Look Once version 5 (YOLOv5) architecture, is proposed for detecting floating feed particles. Firstly, GhostConv, C3Ghost, and Ghost Spatial Pyramid Pooling − Fast (GhostSPPF) modules are used to replace the corresponding modules in the original YOLOv5 backbone and neck. This modification helps reduce the size and computational cost of the model. Secondly, enhancements have been made to the Feature Pyramid Network (FPN) and Path Aggregation Network (PAN) of the original network to improve the recognition of feed. Finally, the space-to-depth (SPD) layer is used instead of the conventional strided convolutions in network architecture. To evaluate the performance of the model, we constructed a dataset focusing on feed particles. The dataset contains 1218 images and 134,640 annotation labels. Experimental results show that the improved YOLOv5s model significantly reduced mis-detections and false-detections produced by other SOTA models, attaining an outstanding [email protected], [email protected]:0.95, precision, and recall rates of 97.0 %, 65.8 %, 98.2 %, 93.3 %, respectively, while maintaining a compact size of 3.3 MB. This approach outperforms newer generalized detection models such as the YOLO12 and offers an efficient means for detecting floating feed particles in aquaculture practices. Additional experiments suggest the great potential of the proposed model in detecting small targets in general.