BSM-YOLO: A Dynamic Sparse Attention-Based Approach for Mousehole Detection

Abstract

In recent years, the proliferation of mousehole in grasslands has exacerbated desertification and compromised grassland productivity, posing potential threats to human safety. Consequently, the identification and forecasting of mouse-hole dynamics for effective infestation control have emerged as pressing concerns. Manual mousehole detection is labor-intensive and time-consuming, hindering comprehensive spatial understanding. Moreover, prevailing detection models lack robust feature extraction for small targets like mousehole, resulting in suboptimal recognition capabilities and diminished accuracy. Addressing these challenges, we propose an enhanced one-stage detection model BSM-YOLO based on YOLOv5 architecture. Firstly, the model integrates a BiFormer module leveraging Bi-Level Routing Attention to capture both global and local features within mousehole images. Subsequently, the incorporation of Shuffle Attention mechanisms enhances the learning of feature dependencies and intricate relationships. Lastly, the adoption of the MPDIoU loss function accurately delineates bounding box characteristics, mitigating redundant box generation and expediting model convergence. In our experimental framework, we curated a dataset comprising 2397 mousehole images to train the BSM-YOLO model. Results indicate that the BSM-YOLO model achieves an average detection accuracy of 94.5%, representing a 5.4% enhancement over the baseline YOLOv5s model. Additionally, the model demonstrates an 8.7 f/s improvement in detection speed. Furthermore, ablation experiments confirm the efficacy of each refinement incorporated into the BSM-YOLO model.

Keywords

• YOLOv5
• object detection
• deep learning
• mousehole