Scalable and transparent automated sewer defect detection using weakly supervised object localization

Deep learning methods for sewer defect detection face challenges due to their reliance on time-consuming bounding box annotations and lack of model interpretability. This paper proposed a framework leveraging weakly supervised object localization (WSOL) that requires only image-level annotations. Analysis showed that effective performance could be achieved with minimal training data (100 images per class) and validation examples (6 images per class). The proposed approach achieved robust performance across six defect classes, with ResNet50 and VGG16 models attaining average MaxBoxAccV2 scores of 64.56 % and 57.33 %, respectively. A two-round evaluation approach was introduced, improving localization accuracy by 10.67 % using ResNet50 backbone. The practical utility of the proposed method was improved through the development of AutoSewerLabeler, a trustworthy prototype tool for automatic bounding box labeling. This paper advances sewer inspection automation by providing a more scalable and transparent framework for defect detection. • CAM-based WSOL framework towards scalable and transparent defect detection. • Achieve reasonable performance with 100 training and 6 validation images per class. • Enable feature-level defect detection using image-level labels. • Develop a trustworthy automatic bounding box labeling tool for sewer defects. • Employ individual optimal threshold for class-specific bounding box generation.