Deep learning-based road crack detection for UAV imagery

UAV aerial imagery-based road crack detection faces three core bottlenecks: high miss rate of tiny cracks, insufficient fine-grained localization accuracy, and excessive parameters hindering UAV edge deployment. To address these, we propose URCD-YOLO, an enhanced small object detection algorithm based on YOLO11s, with targeted optimizations balancing detection accuracy and lightweight design. We optimize the neck network’s bidirectional feature fusion with an improved BiFPN module, replacing conventional concatenation with Softmax-normalized weighted fusion to reduce tiny crack feature loss and fusion deviation during multi-scale aggregation. We build a weighted adaptive dual-path downsampling module (WADown) to replace vanilla convolutional downsampling in the backbone and neck, cutting parameter overhead while preserving micro-crack edge and texture details. Meanwhile, we innovatively embed wavelet transform convolution (WTConv) into the C3k2 module to enhance the network’s perception of low-frequency global road features and fine-grained linear crack textures. We also introduce a multi-scale channel attention (MSCA) mechanism to build a cross-domain collaborative optimization framework covering spatial, channel and frequency dimensions, further boosting the model’s feature capture ability for small cracks in complex backgrounds. Ablation and comparative experiments on the public UAV-PDD2023 dataset show that, compared with the YOLO11s baseline, the proposed algorithm achieves a 9.1% improvement in mAP@0.5 and a 9.5% reduction in parameters, providing a lightweight, high-precision small object detection solution for UAVs and other resource-constrained edge terminals.