A Mix Transformer–Enhanced UNet for Concrete Building Surface Defect Semantic Segmentation

This study proposes mix transformer-enhanced UNet (MiTE-UNet), an improved UNet architecture based on a hybrid convolutional neural network (CNN)–transformer framework, for concrete building surface defect segmentation. The proposed model is designed to address common challenges encountered in practical infrastructure health monitoring applications, including small-object defect recognition, insufficient texture extraction, inconsistency between shallow structural features and deep semantic representations, blurred boundaries, and background noise. To this end, the model integrates the mix transformer encoder of SegFormer to enhance small-object defect segmentation and texture modeling; incorporates channel and spatial attention in the skip connection to unify shallow and deep features; and adopts an improved depthwise separable convolution with channel attention in the decoder to strengthen boundary processing and suppress background interference. The network is provided in three scalable variants—B0, B1, and B2—to balance segmentation accuracy and lightweight design. Experimental results show that MiTE-UNet achieves up to 7.00% improvement in mean pixel accuracy (MPA) and 5.66% in mean intersection over union (mIoU) over its corresponding baseline networks and up to 6.99% and 6.78% improvements in MPA and mIoU over UNet, respectively, exhibiting strong performance and generalization capability. In addition, the proposed model supports real-time inference and can be directly deployed on mobile platforms for real-time detection and alerting of structural defects.