A novel method for silkworm individual recognition under high-density conditions with hybrid module and object detection techniques

Individual detection is a fundamental task for biomass estimation, animal welfare, preliminary disease diagnosis, and intelligent rearing in silkworm culture. However, due to the small size of individuals, occlusion, high density, and complex backgrounds, traditional object detection methods fail to capture global representations, resulting in limited detection performance. To address this issue, we propose a novel method for silkworm individual detection using the You Only Look Once version 5 (YOLOv5) network and a self-attention module. Specifically, multiple local silkworm images were cropped from original images collected using a mobile phone, and labeled using the LabelImg toolkit. The head and tail were used as detection targets to address the challenges of occlusion and high density. A hybrid module was designed to simultaneously extract global and local features by bridging convolutional operations and the self-attention mechanism. An effective detector, termed Hybrid YOLO, was developed by embedding the hybrid module into YOLOv5s. Experimental results demonstrated that the proposed detector achieved a Recall of 91.44% and 81.25%, a Precision of 89.31% and 89.51%, an F1-score of 0.90 and 0.85, and an AP (Average Precision) of 92.87% and 88.98% for the head and tail targets, respectively. The mAP (mean Average Precision) of Hybrid YOLO reached 90.93%, outperforming original YOLOv5 variants, state-of-the-art models, and small target detection networks. This study provides theoretical insights and technical support for future research.