<p>Crack detection is an important prerequisite for evaluating the degree of road damage. Accurate and efficient extraction of pavement crack information from pavement images is the key to the assessment of road technical condition and maintenance condition. The traditional methods (edge detection, morphological method etc.,) are not effective in pavement crack segmentation, the precision of crack contour segmentation is not enough, it is difficult to identify narrow cracks, and the segmentation accuracy is low. Therefore, this paper proposes a novel flexible Transformer network model with ResNet50 for road crack detection. Firstly, a flexible gate is introduced into the self-attention module to select important areas of concern, and a sparse self-attention mechanism from local to global is designed to reduce the computing load and enhance the multi-scale generalization ability of the model. ResNet50 is used as the backbone network to extract pavement crack characteristics. A feature fusion module based on attention mechanism is designed to improve jump connection in ResNet50. Finally, a feature thinning head is added to the decoding part to get an improved model. The experimental results on the pavement crack open data set show that the proposed method can effectively improve the recall rate and accuracy, and has good adaptability to different pavement crack image detection. The accuracy of the proposed method is more than 69%, which is a great improvement compared with most other algorithms. However, very few experimental parameters are required, which aids the computational efficiency.</p>