Abstract: A lightweight improved method based on the YOLO architecture is proposed to solve the problems of excessive model parameters and high computational complexity in road crack detection.The backbone network CSPDarknet53 of the original architecture is replaced with EfficientNet-B0. The average pooling layer and fully connected layers in EfficientNet-B0 are removed to realize the lightweight design of the entire network model. The BiFormer attention mechanism is introduced into the backbone and neck networks to reduce the computational burden of the model. Furthermore, the CARAFE sampling operator is adopted to improve the feature fusion and representation capabilities of the model. Experimental results show that the mean average precision (mAP) of the improved model reaches 88.7%, and the frame rate reaches 86.2 fps. While maintaining a high detection accuracy, the proposed model reduces the network weight, parameter quantity and computational complexity.