LiBiNet: Lightweight boundary information enhanced network for crack segmentation

• Precisely locates cracks and enhances the completeness of crack detection results using boundary information. • The model has fewer parameters, fast inference speed, and can be easily deployed on resource-constrained platforms. • Achieved a successful trade-off between detection accuracy and real-time processing capability. • Outperforms most existing models in detecting small and narrow cracks. Cracks are an essential indicator of infrastructure degradation, such as roads and buildings, and precise pixel-level crack detection is crucial in defect inspection. The current advanced deep-learning networks encounter limitations in simultaneously capturing complete crack structures and achieving high computing performance, leading to incomplete segmentation results and low real-time capability. This hinders the practical applicability in real-world scenarios, especially on mobile platforms. To address these limitations, we propose LiBiNet, a lightweight boundary information-enhanced network. It comprises four main modules: the Boundary Information Module (BIM) enables precise localization and structure extraction of cracks, while the Information Integration Module (IIM) enhances spatial–channel interaction to reduce redundancy and enrich feature representations. The Lightweight Feature Extraction Module (LFM) is designed to efficiently extract deep semantic features, and the Feature Enhancement Module (FEM) further refines the extracted features to boost segmentation accuracy. The evaluation on the CrackVision12K and DeepCrack datasets demonstrates that LiBiNet delivers excellent performance. It achieves F1 scores of 79.36% and 90.48%, mIOU values of 66.34% and 82.69%. For a single 256ⅹ256 image, the inference time is 8.42ms, while requiring only 0.6M parameters and 4.3G FLOPs. These results demonstrate the model’s strong effectiveness and exceptional real-time capability, making it well-suited for practical deployment.