基于改进YOLOv11n的罗汉果炭疽病病害检测研究OA

Siraitia grosvenorii,as an important economic crop with both medicinal and edible values,was severely threatened by anthracnose during its large-scale cultivation.Under the complex field environment,anthracnose lesions were characterized by small scales and blurred boundaries,which led traditional detection models to face the dual challenges of high miss-detection rates and inaccurate localization for small targets.To address this problem,an improved lightweight detection model sensitive to small targets,named WRL-YOLOv11,was proposed.Wavelet Transform Convolution(WTConv)was locally introduced into the backbone of YOLOv11n in this model.With only a slight increase in parameters,the equivalent receptive field was expanded,and the high-frequency responses of boundaries and textures were enhanced,thereby reducing the miss-detection of small targets.A Re-Calibrated Feature Pyramid Network(RCFPN)was designed,and a Selective Boundary Aggregation(SBA)mechanism was embedded in it.This design realized bidirectional re-calibration fusion of shallow and deep features,and a high-resolution P2 layer was added to improve the localization and boundary regression accuracy of small targets.A Lightweight Shared Convolutional Detection Head with Localization Quality Estimation(LSCDH-LQE)was constructed;through explicit localization quality estimation,the classification confidence and localization consistency were improved,and the false positives with high confidence but inaccurate localization were reduced.Experimental results on the self-built Siraitia grosvenorii anthracnose dataset showed that the mAP@0.5 of WRL-YOLOv11 reached 77.0％,which was 5.4％higher than that of the baseline model YOLOv11n;the recall rate was increased to 76.5％,the mAP@0.5:0.95 was increased to 28.9％,the model parameters were 3.43 M,and the GFLOPs were 13.80.The model significantly improved detection performance while maintaining its lightweight nature.Ablation experiments further verified the effectiveness and synergistic effect of each module.In summary,through frequency-domain receptive field expansion,scale re-calibration and localization quality modeling,WRL-YOLOv11 provided an efficient and feasible solution for the accurate detection of small-scale lesions in field environments.