Abstract: The neural network model can automatically identify road defects,however the detection accuracy can not meet the needs of road safety operation and maintenance in practical applications, and it is easy to cause missed detection and false detection of diseases. In response to the above problems, this paper proposes an improved U-Net model that combines grayscale images and depth images.Firstly, the data statistics method based on the depth map is used to automatically eliminate the disease-free data and reduce the computational complexity of the model. Secondly, the global context module is added to the traditional U-Net model structure to realize a lightweight network and improve the network performance on this basis. Finally, the elevation information of the road depth map is added to change the training data of the model from one-dimensional to two-dimensional.Based on the disease range and the pavement depth map, the pavement disease depth parameters are obtained.The results show that the improved U-Net model proposed in this paper, which fuses grayscale images and depth images has global recognition accuracy, accuracy, recall rate, and comprehensive evaluation index and mIoU indicators are 99.09%, 84.69%, 81.64%, 91.67% and 84.58%, respectively, which are higher than the other two models tested at the same time. In the test results of crack disease, This paper based on the improved U-Net model of grayscale image and depth map is 99.07%,which is higher than the other four models.Experiments show that this paper based on the improved U-Net network-based pavement disease identification and extraction algorithm can be used in complex scenes with noise interference. It can extract pavement cracks smoothly and efficiently, and has strong robustness. The algorithm proposed in this paper can provide an important reference for subsequent pavement repair work.

Key words: pavement diseases, 3D point cloud, depth image, improved U-Net model