A Two-Stage Multi-Label Sewer Defect Identification Approach Using TResNet and CSRA Attention Module

To address the coexistence of multiple defect types and class imbalance in complex urban sewer scenarios, this study proposes a two-stage multi-label recognition method for sewer defect images. In the first stage, a lightweight binary convolutional neural network is used to distinguish between normal and defective images, thereby filtering out normal samples and reducing the computational burden. In the second stage, a TResNet-based multi-label classification model is developed, and a class-specific residual attention module is introduced to enhance defect-specific spatial feature representation. In addition, F1-Normal and F2-CIW are used to evaluate the recognition performance from both normal-pipeline recognition and defect-economic-impact perspectives. Experiments on the Sewer-ML dataset show that the proposed method achieves 90.88% F1-Normal and 53.93% F2-CIW, outperforming the baseline models in overall performance.