An Improved Crop Pest Detection Algorithm via Fusing Self-attention and Sample Weighting

In recent years, deep learning has been extensively applied to agricultural pest detection, yet several challenges remain. Variations in object scale, large intra-class shape differences, and high inter-class similarity often lead to missed and false detections. To address these issues, we propose IP-YOLO, a novel detection model that integrates a self-attention mechanism with a sample weighting strategy. Specifically, the CSPBottle module is combined with an improved self-attention mechanism to strengthen the network’s ability to capture positional information of pests. In addition, a cross-layer feature fusion module is incorporated into the Feature Pyramid Network (FPN) to enhance multi-scale feature integration. To further mitigate data imbalance, a sample weighting function is designed to reduce the adverse effects of both easy and hard sample biases on detection results. Experimental evaluations on the Pest36 dataset, collected from real farmland environments, demonstrate that IP-YOLO achieves substantial improvements in detection accuracy compared to existing methods.