YOLOv11-AUC: Improved YOLOv11 for Coal Fly Ash Silo Weld Surface Defect

For the low accuracy of weld surface defects in the coal fly ash silo, an improved YOLOv11-AUC weld surface defect detection method is proposed to avoid omissions and false positives of small defects. The Alterable Kernel Convolution (AKConv) module is introduced into the backbone network to reduce the loss of detail information of feature extraction process, and capture richer channel information. Then, a new cross-scale feature fusion method Upward Integration Feature Pyramid Network (UI-FPN) is proposed to improve the feature representation capability of small-sized object detection. The Context Anchor Attention (CAA) is introduced to the neck network output to enhance the long-distance dependence of the feature map. Finally, the improved algorithm is experimented on the weld defect data set, and the experimental results show that the algorithm achieved an accuracy of 84.1%, the recall rate reaches 76.4%, and the mean Average Precision (mAP) is 83.9%. This study offers an effective solution for weld surface detection of coal fly ash silo.