【Objective】Aiming at the problem of poor real-time performance and low precision of traditional pavement crack detection.【Method】This paper uses the advantages of deep learning network in target detection, and proposes an improved YOLOv5 algorithm, which is called YOLOv5s-attention in this paper, to realize the automatic detection and recognition of pavement cracks. Firstly, the collected crack images are manually labeled with LabelImg annotation software, and then the network model parameters were obtained by improving the YOLOv5 network training. Finally, the model is used to verify and predict the cracks. In addition, F₁ and mAP are used to compare the performance of the original YOLOv5s and YOLOv5s-attention models in detecting and identifying pavement cracks.【Result】The comparison between YOLOv5s and YOLOv5s-attention showed that the precision of YOLOv5s attention increased by 1.0%, F₁ increased by 0.9%, and mAP increased by 1.8%.【Conclusion】It can be seen that the network has certain practical significance in realizing the automatic recognition of road cracks.