Abstract Crack segmentation is crucial for infrastructure maintenance and structural health monitoring. Traditional image processing techniques are limited in their ability to capture complex crack geometries and contextual continuity. Although deep learning methods, particularly Convolutional Neural Networks (CNNs), have made significant advancements in crack detection, they still struggle with long-range dependencies and fine-grained shape capture. To address these challenges, we introduce DFT-CrackNet, a dual-branch encoder-decoder architecture that integrates CNN and Transformer models. Our approach enhances the encoder’s multi-scale feature extraction capability through Hybrid-Pooling and a learnable Linear Perception Guidance Mechanism (LPFM), which emphasizes linear crack structures while minimizing background interference. Additionally, a Multi-Scale Edge Refinement module, based on Double Cross-Centered Differential Convolution (C-CDC), is designed to improve the detection of fine crack edges during the decoding process. DFT-CrackNet is trained on the CrackVision12K dataset and further extended to include fine crack data from tunnels (CrackVision12K-fine), outperforming existing models in complex scenarios. The model achieves superior performance on the CrackVision12K dataset with an F1-score of 82.0%, mIoU of 69.7%, and Recall of 80.9%, and maintains leading performance on the CrackVision12K-fine dataset with an F1-score of 80.9% and mIoU of 68.2%. In this paper, we demonstrate the effectiveness of our model in accurately segmenting cracks, particularly in fine-grained and low-contrast environments, showcasing its significant potential for real-world infrastructure maintenance applications. The source code is publicly available at https://github.com/li5jing/DFT-CrackNet.
Feng et al. (Thu,) studied this question.