# DCP-TransUNet: An Approach for Crack Segmentation on Roads

## TL;DR

This paper presents DCP-TransUNet, a new model for identifying cracks in roads, which performs well on both public and private datasets.

## Contribution

The novel DCP-TransUNet model combines a hybrid encoder with new modules for improved crack segmentation in complex road conditions.

## Key findings

- DCP-TransUNet achieves 79.12% mIoU and 87.96% Recall on a public crack dataset.
- On a private dataset, it reaches 68.83% mIoU and 81.67% Precision, showing strong adaptability.
- The model outperforms existing methods in crack segmentation accuracy and interpretability.

## Abstract

For cement pavements on vast road networks, cracking has become one of the principal distresses threatening structural integrity and traffic safety. This study introduces DCP-TransUNet, a model featuring a new hybrid encoder that enhances the continuity of crack extraction under complex conditions through a DSE-CNN module and a CLMA-Transformer block. To further strengthen learning and interpretability for challenging crack imagery, a PPA bottleneck module is designed to capture additional discriminative features. Experimental results indicate strong performance: on the public dataset, DCP-TransUNet achieves mIoU 79.12%, Recall 87.96%, F1 87.06%, and Precision 86.21%; on the private dataset, it attains mIoU 68.83%, Recall 74.42%, F1 77.57%, and Precision 81.67%. Compared with other models, these outcomes demonstrate the method’s accuracy and effectiveness for crack segmentation.