MFDF-YOLO:复杂场景下的轻量级棉花检测算法OA

To address the challenges of false detections,missed detections,and inaccurate localization in complex cotton field environments characterized by densely distributed,multi-scale cotton targets,background interference,and occlusion,this paper proposes a lightweight cotton detection algorithm based on YOLOv11,named MFDF-YOLO.The algorithm introduces a multi-scale edge feature selector(C3k2-MSEFS)to replace the original C3k2 module in the backbone,which enhances high-frequency edge features and suppresses background noise through edge feature augmentation and dual-domain selection,improving the model's ability to perceive and localize multi-scale targets.Additionally,a context anchor attention(CAA)-integrated hierarchical scale-based feature pyramid network(HSFPN)is designed to restructure the neck network,applying spatial dynamic reweighting and selective multi-scale feature fusion to maintain the response of small and occluded targets during feature aggregation.Finally,a lightweight efficient detection head(LEDH)is devel-oped using grouped convolutions and a parameter-sharing structure,significantly reducing computational cost while en-hancing detection accuracy.Experimental results on the self-built cotton dataset show that MFDF-YOLO improves the mAP@0.5 by 5.2 percentage points compared to the baseline model,reduces the parameters by 30.6％,the computational cost by 12.7％,and the model size by 24.1％.Additionally,the MFDF-YOLO model is verified by the COCO and TIDE metrics to exhibit significant advantages in multi-scale object detection,localization capability,and background suppres-sion,and its favorable generalization ability is further validated on public datasets.