Research on Road Pothole Detection and Area Estimation Based on Transfer Learning and U-Net Model

To address the issue of inadequate accuracy in traditional target detection and edge feature extraction algorithms for pothole detection on roads, this paper proposes a pothole detection and area estimation method based on transfer learning and the U-Net model. The method employs a pre-trained ResNet50 transfer learning model for pothole recognition and integrates the U-Net semantic segmentation model to achieve precise edge detection and area estimation of potholes. To enhance the model's performance, various data augmentation techniques were applied to the training dataset. Through comprehensive multi-dimensional evaluation, the recognition rate of the transfer learning model reached 95%, demonstrating excellent robustness and generalization capabilities. Furthermore, the U-Net model, utilizing the pre-trained ResNet50 as the backbone feature extraction network, performed outstandingly in pothole area estimation, achieving a mean Intersection over Union (mIOU) and mean Pixel Accuracy (mPA) of 91.67% and 94.99%, respectively. In conclusion, the proposed method offers significant application value in pothole recognition, edge detection, and area estimation on roads.