Hybrid Multi-Scale CNN-Transformer Network for Structural Surface Crack Segmentation

• Proposed CrackHCT-Net, a hybrid CNN-Transformer for structural surface crack segmentation • Proposed an IDSC-based gated linear unit to further enhance contextual feature representation • Proposed an efficient feature fusion module to fuse local and global feature information • Validated the model on three public datasets to evaluate the model’s performance • Performed ablation studies to verify the effectiveness of each component Cracks are surface-level structural defects commonly found in built infrastructure at various scales and types, with distinct edges and textures. Additionally, the proportion of the crack surface is extremely small compared to the background surface, making accurate detection is challenging under diverse structural background conditions. Regular monitoring is therefore essential to maintain structural integrity and safety. This task requires a model capable of detecting cracks of varying shapes and backgrounds while remaining computationally efficient for automated anomaly detection. To address this, we propose CrackHCT-Net, a hybrid multi-scale network that combines Convolutional Neural Network (CNN) and Transformer architectures for crack segmentation. The model employs a lightweight CNN encoder to extract local features and a Transformer encoder incorporating lightweight attention and inverted depthwise separable convolution-based gated linear units to capture discriminative global contextual information. A multi-scale feature fusion module is introduced to aggregate features extracted by the CNN and Transformer encoders at the same semantic level while minimizing discrepancies in their feature representations, and reducing redundant features extracted by both encoders. Experiments on three public crack datasets: Crack3238, DeepCrack537, and CFD demonstrate that our method achieves strong segmentation performance, with IoU scores of 63.31%, 76.99%, and 58.33%, respectively.