Abstract: Addressing the current challenges of low detection accuracy and missed detections due to complex backgrounds and small sizes of these pests, an enhanced YOLOv5s method for detecting stored grain pests is proposed. Firstly, MobileNetv3 is employed as the backbone feature extraction network for the YOLOv5s model, with its SE attention mechanism being modified into an ECA module to reduce computation and parameter count. Simultaneously, the PANet network is adjusted to incorporate a weighted bidirectional feature pyramid BiFPN structure to enhance feature fusion capabilities. Secondly, Swin Transformer is introduced into the neck network to address insufficient global feature extraction issues and improve recognition accuracy. Finally, the EIOU loss function is utilized to expedite model convergence speed. The results demonstrate that this improved model achieves a mAP (mean average precision) score of 97.8% and operates at a FPS (frames per second) rate of 133.3, showcasing remarkable robustness and generalization abilities when compared with mainstream target detection models. This study's improved YOLOv5 model for detecting stored grain pests effectively overcomes environmental complexities while significantly enhancing pest detection under conditions involving dense distributions and occlusions-thereby providing valuable technical insights towards achieving real-time pest detection on mobile devices.