LightCSeg: Lightweight Crack Segmentation Network With Adaptive Sobel And Local Enhancement

Real-time crack segmentation is a key technology for infrastructure health monitoring. Existing methods fail to address the challenges of lightweight design and practical applicability. To tackle this challenge, we propose a lightweight real-time crack segmentation network, LightCSeg. First, we designed the Sobel-Guided Adaptive Edge Aware (SAE) module, which fully leverages the edge feature extraction capability of the Sobel operator and the adaptive capacity of convolutional layers. Through a simple yet effective design, it significantly enhances the response intensity of crack textures. Second, we designed the Local Enhanced Multi-Scale Fusion (LEF) module, which employs bilinear interpolation for upsampling while leveraging the advantages of neighborhood attention for local modeling, ensuring high-quality feature fusion. Experiments on benchmark datasets for crack segmentation demonstrate that our method achieves SOTA-comparable performance while significantly reducing both parameters and FLOPs.