Enhanced concrete crack detection and proactive safety warning based on I-ST-UNet model

Existing Swin-Transformer-UNet models for concrete crack detection have several limitations, such as weak feature extraction and loss of detailed image information. This paper thus presents an improved Swin-Transformer-UNet (I-ST-UNet) model by embedding Swin-Transformer into UNet. An experimental dataset with 2030 images was used to train and test the I-ST-UNet model. The model showed significant segmentation performance improvements: 0.7% in Accuracy, 2.25% in Mean Accuracy, 5.77% in Mean Intersection-Over-Union, and 1% in Frequency Weight Intersection-Over-Union. Relative errors for crack widths from 0.1 mm to 0.2 mm and those exceeding 0.2 mm consistently remained below 5%. For crack widths over 0.2 mm, the model achieved a remarkable 98.35% accuracy in safety warning. Important features of the I-ST-UNet model include (i) adding External Attention to the skip connection layer to enhance the potential relationship between samples; (ii) improving Feature Pyramid Network to condense features at different scales during upsampling in the decoding phase. •Development of an improved Swin-Transformer-UNet (I-ST-UNet) model.•Integration External Attention into the skip connection layer of the I-SI-UNet model.•An improved Feature Pyramid Network for the decoding process of the I-ST-UNet model.•Accurate identification concrete cracks using the I-ST-UNet model.•Successfully achieving concrete crack width calculation and safety warning.