Pavement crack is the most common defect of road. With the development of deep learning technology, more and more methods are used to extract the crack information from pavement images. Aiming at the problems of low accuracy and lack of real-time due to incomplete extraction of crack features by existing deep learning pavement crack detection methods, a road crack detection method combining attention mechanism and GhostUNet is proposed. This method is composed of encoder and decoder. The conventional convolution in U-Net is improved to Ghost convolution and the number of model parameters is reduced. In coding and decoding, in order to improve the ability to extract crack features, ECA attention mechanism and residual connection are introduced. ECA attention module can filter irrelevant feature information, and residual connection can be used to avoid network degradation. To evaluate the effectiveness of this method in fracture detection, two publicly available fracture data sets were used, and ablation and comparison experiments were conducted. The experimental results of F1_score, P and R increased by 14.48%, 14.35% and 14.45%, respectively, compared with U-Net. The number of parameters in this model decreased by 14.2 MB compared with U-Net. Compared with similar models, this model has higher segmentation accuracy and fewer parameters.