Cas-SegFormer: A Concrete Crack Semantic Segmentation Model Based on Convolutional Additive Self-Attention

To address the issues of low efficiency in resourceconstrained scenarios, insufficient feature utilization, and multilevel semantic degradation (information loss limiting detail recovery) in existing concrete crack detection models, the lightweight Cas-SegFormer model is proposed. This model introduces a Channel-Priority Convolutional Additive Self-Attention(CPCAS) Mechanism to enhance the model's perception of channel and spatial information, thereby improving its global modeling capability. While ensuring model efficiency, it significantly reduces the number of parameters. Additionally, a Wavelet-based Subpixel Fusion Module (WSFM) is designed, incorporating wavelet convolution to enhance the model's perception of frequency information. The sub-pixel convolution module is employed for pixel rearrangement, refining the encoder's output and achieving upsampling with less information loss. To further improve crack detection accuracy, a hybrid loss function combining Lovász and Dice is proposed. The Lovász loss directly optimizes IoU to refine boundary detection, while the Dice loss optimizes overall segmentation consistency. By dynamically weighting and combining the strengths of both, the model can mitigate class imbalance issues and enhance its ability to capture fine structures such as cracks. Experiments on a mixed public dataset show that the improved SegFormer model achieves a 9.61 % increase in mIoU compared to the original SegFormer-b0. In terms of the$\text{mF1}$metric, it outperforms mainstream semantic segmentation models such as DeepLabV3+,Fast-scnn,and Segmenter by 2.64 %, 5.1 %, and 3.27 %, respectively, while maintaining a computational cost of only 0.67 GFlops.