Research on the Real-time Monitoring System of Mouse Granary Based on the Improved YOLOv5

In agriculture, preventing rodents from invading grain bins has always been an important task. However, traditional grain bin monitoring systems are often unable to issue timely and accurate alerts. Therefore, this study proposes a mouse detection method based on the improved YOLOv5s algorithm, which aims to protect grain bins from mice. The method is mainly based on replacing part of the C3 module of the YOLOv5 backbone network in order to reduce the network parameters and computational complexity, and to improve the inference speed and efficiency of the model. At the same time, a $160 \times 160$ scale small target detection layer is added to improve the extraction capability and detection accuracy of the features of the long-distance mouse targets. In addition, Swin-Transformer technology is introduced to better capture long-distance relationships using the selfattention mechanism to improve the effect of image classification and target detection. The CBAM attention mechanism is also introduced to enhance the robustness and performance of the network in complex environments when detecting mice. The boundary loss function Elou is adopted, which helps the model predict the mouse boundary position more accurately and improves the stability and training speed of the model. The experimental results show that the improved YOLOv5s model improves in precision, recall and mAP values, from the original $\mathbf{9 0. 2 \%, 8 7. 5 \%}$ and $\mathbf{8 8. 5 \%}$ to $\mathbf{9 3. 8 \%, 9 0. 7 \%}$ and $\mathbf{9 2. 5 \%}$, and the average detection time is further reduced from 28 ms to 26 ms.