A Multi-scale Features Fusion Method for Concrete Crack Segmentation

Cracks are a major defect in concrete bridges, and accurate segmentation is crucial for their detection. The complex topology, low contrast, and noisy background make crack detection discontinuous and insufficient in detecting small branches. Thus, crack detection remains a challenging task. This paper proposes a model that integrates semantic and detail features, specifically designed for multi-branch cracks. The model consists of two components: the high-level semantic extraction part and the shallow detail fusion part. The semantic extraction part is based on the U-Net, and optimized with the Efficient Channel Attention (ECA). We introduce a Max-pooling Efficient Channel Attention Block between the encoder and decoder to aggregate larger spatial information and preserve more texture information of cracks. Subsequently, we establish the detail part to enhance the detection accuracy of small branches. This is achieved by connecting shallow multi-scale features, which retain important detail information from different global morphologies, thereby better constraining the topological continuity of the segmentation. Experimental results on our custom dataset, our method has surpassed the state-of-the-art TransUNet model. Compared to the baseline model. Our approach achieved improvements of 0.009 and 0.147 in F1 scores on both datasets.