Abstract:

Road pavements, as critical transportation infrastructure, are prone to cracks and other defects due to long-term load bearing and environmental erosion. Traditional manual inspection methods suffer from

low efficiency, high subjectivity, and significant safety risks, failing to meet the maintenance demands of large-scale pavement networks. This paper proposes LMC-YOLO (Lightweight MobileNetV4 with CAA for YOLO), a lightweight and high-precision pavement crack detection algorithm that addresses feature loss and insufficient accuracy in detecting thin and elongated cracks. The algorithm systematically optimizes the backbone structure, attention mechanisms, and lightweighting strategies of the detection network. By incorporating the lightweight MobileNetV4 with an efficient Universal Inverted Bottleneck (UIB) architecture as the backbone, a balance between powerful feature extraction and low computational overhead is achieved. The improved Context Anchor Attention (CAA) mechanism is integrated into the Neck of the detection network, introducing a shape-aware module based on strip convolution that effectively enhances detection capability for thin elongated cracks. Through refined network design, the model reduces parameter count by 24% and achieves a computational cost of 4.6 GFLOPs. Experimental results demonstrate that LMC-YOLO achieves 91.1% precision, 84.6% mAP@0.5, and 70.3% mAP@0.5:0.95 in crack detection tasks, with an F1-score of 80.40% and an inference speed of 345 FPS. Cross-dataset validation on DIOR and DOTA-v2 further confirms the model's cross-domain transfer capability. This method successfully combines high accuracy with efficient lightweighting, providing a practical solution for real-time pavement crack detection on mobile and embedded devices.

Key words: road , engineering, crack detection, lightweight model, attention mechanism, deep learning