A multi-label classification method for sewer defects based on the collaboration of local and global features

Sewer defect detection plays a crucial role in ensuring the efficient operation of urban sewage treatment infrastructure. However, most existing sewer defect classification models rely on generic convolutional neural network architectures, which often overlook the inherent characteristics of sewer defects and limit their performance in multi-label scenarios. To address this issue, this paper proposes a dual-backbone network that integrates local and global feature extraction for automated sewer inspection. First, the network employs ResNet18 to extract detailed local features from sewer images, capturing texture and structural information, while a lightweight vision Transformer models global contextual dependencies in complex scenes. Simultaneously, the Adaptive Dual-Stream Feature Alignment Module (ADSFAM) dynamically adjusts and fuses the features from the two backbones, significantly enhancing the consistency and discriminative ability of the combined features. Second, a global-local attention mechanism in the decoder reinforces the interaction between fine-grained local details and global semantics. Furthermore, an Asymmetric Loss function is adopted to address the class imbalance issue inherent in multi-label tasks. Extensive experiments on the Sewer-ML dataset demonstrate that the proposed method achieves an F1 score of 91.48 and an F2_CIW score of 67.10, outperforming existing approaches while maintaining high inference efficiency.