Abstract <p>Regular inspection of underground sewer networks is essential for the healthy operation of drainage systems and urban safety. This paper proposes a framework based on an enhanced Transformer for small-sample defect classification and localization. The framework includes data augmentation using Mixup, a novel Feature-aware Swin Transformer (FSwin_T), and class activation mapping (CAM) for weakly supervised localization. The proposed FSwin_T model combines the features of the residual module and the Transformer, increases the complexity of extracting feature information, and solves the poor training effects on small-sample datasets. We address the issue of poor classification accuracy in small-sample datasets by using Mixup and employ five CAM techniques to interpret the FSwin_T model, enabling defect localization and visualization. Our research demonstrates that the proposed enhanced Transformer framework has better defect classification results without relying on manual annotation, locates and visualizes key information in images, and provides certain explanations for inspectors to identify pipe defects.</p> Graphical abstract <p></p>