Research on fine-grained segmentation algorithm based on transformer-UNet hybrid architecture

As the aging of global infrastructure intensifies, detecting cracks has emerged as a crucial task for maintaining the structural health of infrastructure and public facilities.This paper proposes a dual-path Transformer-UNet hybrid architecture (DPT-UNet), which integrates global context modeling and local feature extraction to solve the challenge of crack segmentation in complex scenes. DPT-UNet employs a dual-path attention mechanism (DPA) to handle both global self-attention and local convolutional features simultaneously, using dynamic gating weights to enable the adaptive integration of features. It also combines a dynamic feature selection module (DFS) to optimize the allocation of multi-scale feature weights. By using knowledge distillation technology, the model parameters are compressed to 2.1M, achieving real-time inference on mobile devices (32 FPS). Experiments on the CrackTree200 and self-built BridgeCrack datasets show that the mIoU of DPT-UNet reaches 89.7%, which is a 6.2% improvement over the best existing models, and the parameter volume is reduced by 83%. Compared with Swin-UNet, our model has a significant advantage in the balance between parameter volume and accuracy. Additionally, the cross-dataset validation results on the Crack500 dataset further prove its generalization ability. This research offers an effective approach to fine-grained crack segmentation and facilitates the deployment of lightweight models in industrial inspection.