Ceramic Ring Defect Detection: An Algorithm Constructed on the Improved YOLO11 Model

To achieve efficient detection of ceramic ring defects, a lightweight YOLO11-SDE defect detection algorithm is proposed. Firstly, in order to enhance the focus on feature detection and improve computational efficiency, the C3k2-DilResNet, which combines dilated reparametrization blocks and dilated residual designs, is used to replace the second, fourth, sixth, and eighth layer C2f modules in the original YOLO11’s backbone. Secondly, to address the difficulty in extracting defect features, a spatial pyramid pooling combined with a local spatial attention mechanism is formed in the ninth layer $\mathbf{C 2 f}$ module of the original YOLO11 backbone, resulting in the SPPF-LSKA module. Then, to further improve the lightweight design and feature perception, the CAA-HSFPN module and the added ConvTranspose2d convolution are used to construct a new lightweight Neck to replace the original Neck in YOLO11. Finally, experiments are conducted on a self-made dataset, showing that the improved algorithm has reduced computational requirements, parameter count, and weight file size compared to the original YOLO11 model, with a 1.8% increase in mAP, achieving $90.3 \% \mathrm{mAP}$ for various defects in ceramic rings; the proposed algorithm requires fewer computational resources while maintaining higher detection accuracy.