| Concrete cracks pose a significant safety hazard to concrete structures. To effectively segment concrete cracks while promptly detecting and evaluating their potential risks, proposes an improved lightweight crack semantic segmentation algorithm based on DeepLabV3 . Firstly, MobileNetV3 is employed as a lightweight backbone network to significantly reduce the model's parameter size. Secondly, the Attention-based Intrascale Feature Interaction module (AIFI), which integrates feature interaction within scales, replaces the dilated convolution branch, and the Normalization-based Attention Module (NAM), is introduced to enhance the global information representation of features, and promoting multi-level interaction of crack feature information. Additionally, the mixed model of both self-Attention and Convolution (ACmix) , is introduced after extracting low-level features to more effectively capture detailed features of high-resolution crack images. Finally, the C2f-SCConv module, combining Spatial and Channel Reconstruction Convolution (SCConv) and Global Attention Mechanism (GAM), is proposed to decode the fused high and low-level feature streams, reducing computational redundancy while improving the model's perception of multi-scale semantic and detail information. Experimental results compared with other segmentation algorithms on the public crack datasets, Crack3k and Asphalt3k, demonstrate that the proposed concrete crack detection model reduces the parameter size by 88.1% and the floating-point operations by 83.8% compared to the DeepLabV3 model. It achieves a pixel accuracy improvement of 0.02% and an average intersection over union of 86.21%, The average frame rate is 47.91 frames per second. The model significantly improves segmentation performance while reducing model complexity. |