By integrating GAN (Generative Adversarial Network) for image generation and enhancement, and improving the U-Net network with the Convolutional Block Attention Module (CBAM), a high-precision recognition model for drainage pipeline defects is proposed to accurately and automatically segment and identify the image defects of drainage pipes. The model achieves satisfactory recognition results. The mPA (Mean Pixel Accuracy) is 91.40% and the mIoU (Mean Intersection over Union) is 85.02% on the validation set. The results indicate that the generation and enhancement of images with fewer defect categories in the collected data by using GAN can greatly improve the accuracy of defect recognition with fewer samples. After data augmentation, the DeepLabv3+ networks of PSPNet, U-Net and different backbone networks have been significantly improved, with an average increase of 20%~30% in mPA and mIoU. Adding the CBAM attention module can enhance the segmentation ability of U-Net and enable the model to have better pixel-level segmentation ability.