Abstract: Cracks on concrete surfaces in underground engineering are one of the key factors affecting construction safety. Accurate and efficient crack detection can help prevent safety incidents to a certain extent. To address this issue, an improved YOLOv8 model for cracks on concrete surfaces detection and identification is proposed. Firstly, the backbone network of YOLOv8 is enhanced by incorporating dilated convolutions to improve feature extraction for shallow-layer targets. Secondly, the  CBAM (convolutional block attention module) is introduced to strengthen the model’s ability to capture crack features. Thirdly, the neck network structure of YOLOv8 is modified to address the challenges of small target features being too miniature and weak feature textures making learning difficult. Finally, the feature fusion method in the neck network is optimized. Experimental results show that the improved YOLOv8 model achieves a 36.94% increase in Precision, a 49.18% increase in Recall, and a 51.74% increase in mAP(mean average precision). The enhanced model is better suited for concrete crack detection in complex scenarios and further improves the recognition performance for small targets in challenging environments.

Key words: convolutional network, attention mechanism, feature fusion, network optimization, YOLOv8