CDNetYOLO: a multidefect detection model for concrete walls integrating dynamic convolution and large kernel attention

To address the challenges of concrete wall defect detection—such as large surface areas, small target sizes, low contrast, and background clutter—this paper presents CDNetYOLO, an improved YOLOv11n-based detection framework. The backbone integrates KWConv to enhance feature adaptability while maintaining computational efficiency. A modified SPPF module, termed SPPF_LSKA, incorporates Large-Kernel Self-Attention to better capture small defects and suppress noise. Additionally, a novel neck structure, GDFPN, combines RepGFPN and DySample to strengthen multi-scale semantic fusion and spatial alignment. Experiments on a custom dataset with six defect types demonstrate that CDNetYOLO achieves 70.6% [email protected] and 50.0% [email protected]:0.95, offering an effective balance between precision and efficiency for real-time industrial scenarios, while maintaining a low computational complexity of 6.2 GFLOPs.