Building concrete surface crack detection method based on improved YOLOv8

This article proposes a method for detecting surface cracks in building concrete based on improved YOLOv8. By introducing deformable attention mechanism (DAttention) in the backbone network, the crack growth trend can be dynamically focused; Enhance multi-scale feature expression capability in neck design (Cross scale Feature Fusion Module, CCFM); Embedding Efficient Channel Attention (ECA) in the head to enhance the weight of key features; And replace the Complete Intersection over Union (CIoU) loss function with the Scale invariant Intersection over Union (SIoU) loss function to optimize the bounding box regression process. The experimental results show that our method achieved a detection accuracy of 88.4%, a recall rate of 95.2%, and an average precision mean (mAP) of 96.4% on a self built dataset, which is significantly improved compared to the benchmark YOLOv8 model. This method is capable of extracting crack features in a complete and continuous manner, effectively identifying subtle cracks and suppressing background interference, providing reliable technical support for the health assessment and safety risk prevention of building structures.