DMF2Net: dual-branch multistage feature fusion network for road crack segmentation

Road crack segmentation presents several challenges, including variations in background, complex crack textures, and high levels of image noise. Existing crack segmentation algorithms often fall short in accuracy and robustness. To address these issues, this study introduces a dual-branch multi-stage feature fusion network (DMF2Net) for road crack segmentation. Specifically, In the one branch, Swin Transformer is used as a backbone architecture to extract global features. While the local information is extracted by a convolutional neural network (in which the residual network is used as a backbone network) in the other branch. The same scaled global and local information is fused by a fusion module combines the features from both encoders, while a cross-stage fusion structure is employed to enhance detail extraction and minimize class inconsistency. By fully utilizing the multi-scale characteristics of Swin-T and CNN, DMF2Net achieves greater accuracy and robustness than existing models. To validate the network’s effectiveness, comprehensive experiments were conducted on three publicly available crack datasets. The results demonstrate that DMF2Net outperforms state-of-the-art crack detection methods across all six evaluation metrics.