Self-Derived Multimodal Knowledge Distillation for Real-Time Road Crack Detection

Road crack detection is vital for intelligent transportation and infrastructure maintenance. In recent years, deep learning-based methods have emerged for automated road crack detection with enhanced performance. Nevertheless, existing methods fail to balance detection performance, model complexity, and processing efficiency, hindering their deployment in autonomous road inspection systems. To fill this gap, we propose a novel lightweight network for real-time road crack detection, called LR-CrackNet. It employs a teacher-student network architecture, where the teacher network leverages both image data and adaptively self-derived boundary information from pixel-level annotations, while the student network relies solely on image input. The training process focuses on distilling the teacher network’s detection capabilities into the student network. Both networks share a U-shaped segmentation design, integrating novel residual dual-layer depthwise convolution blocks and a fast Transformer block for efficient integration of detailed hierarchical spatial information and global long-range contextual information. Furthermore, an attention-enhanced discriminator network is designed, which aims to improve detection performance by enforcing global consistency and providing pixel-level feedback. Comprehensive experimental results demonstrate that LR-CrackNet achieves state-of-the-art (SoTA) detection performance on the UDTIRI-Crack, DeepCrack, and CamCrack789 datasets and surpasses existing publicly available algorithms in processing efficiency, with 324.40 frames per second (FPS) on a single NVIDIA RTX3090. The source code package is available at https://mias.group/LR-CrackNet/.