Lightweight pavement crack detection model for edge computing devices

In response to the issues of false detection, missed detection, and large model parameter volume during the detection of road surface cracks in complex backgrounds, a lightweight road crack detection model named YOLO-DGVG based on deformable convolution is proposed. Firstly, deformable convolution DCNv2 is introduced into the backbone network, and the C2f-DCNv2 structure is designed to enhance the network's adaptive adjustment capability to the shape of road surface cracks. Secondly, the lightweight convolution technology GSConv is introduced into the neck network to replace the Conv layer in the neck network for feature extraction, and the original C2f module is replaced with the VoVGSCSP module, which improves the detection accuracy of the model while reducing computational complexity. Finally, a grouped convolution detection head module GCH is constructed in the head network to further reduce the model's parameter volume. To verify the effectiveness of the improved parts, experiments were conducted using the PID dataset for ablation studies compared with the YOLOv8 model. The Recall and mAP were increased by 0.3-1.6%, respectively, while the Para was reduced by 22.28%. Additionally, generalization experiments were carried out using the UAPD, RDD2022, and self-built datasets. To further validate the overall effectiveness of the model, comparisons were made with models such as RT-DETR and YOLOv10, with YOLO-DGVG outperforming the comparison models. The model was also deployed on edge computing devices to achieve crack detection in static road surface images.