HSFL-YOLO: Improved YOLOv12-based target detection of wheat aphids in the field

Wheat aphids are major pests threatening wheat production, and accurate field detection is essential for pest management and food security. However, in complex field environments, aphid detection remains challenging due to the small size of aphids, weak visual features, and severe background interference, which often result in missed detections and false alarms in traditional methods. To address these issues, this study proposes a YOLO (You Only Look Once)-based small object detection method called HSFL-YOLO. First, a HAFB (Hierarchical Attention Fusion Block) feature fusion module is designed to enhance information exchange among multi-scale features and improve the fusion of shallow detail features with deep semantic features. Second, the original PAFPN (Path Aggregation Feature Pyramid Network) is optimized by constructing a SOEP (Small Object Enhancement Pyramid) structure, which retains more positional information for small targets and enhances the model’s perception of small objects. In addition, the FRFN (Feature Refinement Feed-Forward Network) module is introduced to adaptively reconstruct features and strengthen channel-wise interaction, thereby improving feature representation in dense small-target regions. Finally, the GCD(Gaussian Combined Distance) loss and Focal-EIoU(Focal Efficient Intersection over Union)loss are combined to enhance the model’s focus on hard samples and improve bounding box regression accuracy. Experimental results on the wheat aphid dataset show that, compared with the baseline model, HSFL-YOLO improves precision, recall, mAP(mean Average Precision)@0.5, and mAP@0.5:0.95 by 2.3%, 1.5%, 0.4%, and 1.1%, respectively. The proposed method effectively improves small object detection performance while maintaining reasonable model complexity, providing technical support for intelligent pest detection in complex agricultural scenarios.