Abstract
This article proposes an improved model based on UNet to adapt and segment cracks in asphalt pavement. The model uses ResNet50 as the backbone feature extraction network to solve the problem of degradation during deep neural network training. To address the issue of uneven distribution of positive and negative samples caused by the small proportion of crack pixels in the image, a mixed loss function is used to balance the loss caused by positive and negative samples. In response to the useless feature information generated by concatenating the convolutional feature map in the encoder with the corresponding upsampled feature map in the decoder, a CBAM attention mechanism module is introduced after concatenation, and at the same time, a CBAM attention mechanism module is introduced after upsampling in the decoder, allowing the network to further learn the defect feature information, ignoring the unimportant information in the input image, and improving the learning efficiency of the model. The experiment shows that the improved UNet model has improved Recall, MIou, and mPA values compared to the original UNet model, and its performance is better than other classic semantic segmentation models. The segmentation effect is also improved, which is beneficial for crack detection and subsequent road maintenance