YOLO-CESn: an efficient field weed detection model based on an improved YOLO11

Abstract Weeds in farmland severely reduce crop yield and quality. However, traditional detection models struggle with weed diversity, blurred boundaries, limited multi-scale feature extraction, and occlusions. To address these challenges, this study proposes an improved detection model, YOLO-CESn, based on You Only Look Once 11 n (YOLO11n). First, deformable convolution v4 (DCNv4) and ghost modules are integrated within a cross stage partial with kernel size 2 (C3k2) structure in the backbone to enhance geometric feature extraction and reduce false positives caused by diverse weed morphologies and unclear boundaries. Second, an efficient multi-scale fusion neck is designed by combining low-level and high-level features with a tiny object detection head, thereby improving recognition of early-stage weeds and achieving full coverage across growth stages. Finally, a soft non-maximum suppression post-processing mechanism is adopted to decay confidence scores of overlapping bounding boxes instead of applying hard suppression, thus alleviating missed detections under dense distributions and occlusion. Experimental results show that for the Fine24 dataset, YOLO-CESn achieves 74.0% [email protected] and 52.1% [email protected]:0.95, representing improvements of 4.5% and 3.7% over YOLO11n, respectively. For the CottonWeedDet12 dataset, the model attains 94.0% [email protected] and 88.4% [email protected]:0.95, with corresponding increases of 1.8% and 1.6%. With only 7.4M parameters and 120 FPS inference, YOLO-CESn provides a lightweight and effective solution for weed detection in precision agriculture.