Crop pest detection method based on improved YOLO11

In view of many problems such as the complex background of pest detection, the wide variety of species, and the small size of some pest species that are difficult to accurately detect, this paper studies a high-precision detection algorithm (HFT-YOLO) based on YOLO11. In the HFT-YOLO algorithm, Token Statistics Self-Attention is used to improve C2PSA to design the CSPTSSA module, which not only curtails the computational complexity inherent in the model but also significantly bolsters the interpretability of the model; in addition, Adaptive Fine-Grained Channel Attention is used to improve the C3k2 module in the backbone network to design the C3k2_FCA module, which improves the accuracy of feature weight allocation by the model; finally, the HyperC2Net network structure and ADown module are used to redesign the HCANet module for the YOLO11 neck structure, allowing complex high-order interactions between levels and positions and improving the neck feature extraction capability. Experimental findings demonstrate that the detection mean average precision atmAP50 and mAP95 of the refined model attain 83.5% and 53% respectively, which are 2.1 and 1.7 percentage points higher in comparison to the original model, effectively augmenting the model's accuracy.