Aiming at the problem of small amount of original fault data and unbalanced data set of bearing rollers in actual production, a data enhancement strategy was proposed to expand the original bearing image data set, and combined with the U-Net framework and lightweight deep learning model to construct an end-to-end bearing roller semantic segmentation model method. By combining the U-Net framework and lightweight deep learning models MobileNetV1 and DenseNet121, the end-to-end bearing roller semantic segmentation models LS-MobileNetV1 and LS-DenseNet121 are constructed,the proposed models are trained based on the transfer learning strategy, and compared with other models for experimental analysis. The results show that compared with the existing methods, the method in this paper achieves higher segmentation accuracy and more robust detection results with few parameters, which verifies the effectiveness of the proposed method.