Abstract: Currently, there is a scarcity of road disease detection data sets, single detection scenarios, and the existing road disease detection methods based on deep learning are difficult to deal with complex environmental interference, and the model size is too large to deploy. A multi-type and scenario oriented pavement disease detection data set was established to make up for the shortcomings of existing data sets. Furthermore, a pavement disease detection method based on improved YOLOv5 was proposed. This method integrated an attention mechanism and lightweight structural components to improve the model detection accuracy while reducing the number of parameters, achieving the detection and accurate identification of cracks and potholes pavement damage under various interference backgrounds, and effectively improving the aforementioned deficiencies. Results show that the proposed method has a high detection accuracy of 93.3% on the constructed pavement disease data set, and the number of model parameters is only about 6.7 × 10⁶, which greatly reduces the deployment cost.