| description abstract | In the field of sewage pipeline inspection, the Pipeline Assessment Certification Program (PACP) requires encoding pipe cracks at the subcategory level in practical applications. This study seeks to address the need to improve the accuracy of the crack pixel-level segmentation and subcategory classification problem based on sewer pipe inspection video images. The proposed framework consists of a feature extraction module, a multiscale feature fusion module, a mask-guided attention module, and a subcategory classification module. First, the multiscale feature fusion mechanism is used to integrate the convolution features of different scales to obtain the convolution features rich in both semantic and detailed information, which can improve the characterization ability of cracks and achieve accurate crack detection. Then, mask-guided attention is used to refine features and eliminate background noise, enhancing the feature representation for crack pixels and reducing the ambiguity of the fine-grained crack classification task as well. Finally, the subcategory classification result is obtained based on the feature maps with erased background clutter. The experimental results show that the proposed method achieved F1-scores of 97%, 94%, and 95% for longitudinal, circumferential, and multiple cracks, respectively. The main contribution of the study is the formulation of a multitask collaborative learning framework, which combines pixel-level segmentation and global image-level classification annotation to achieve a desirable performance of sewer crack detection and classification at a subcategory level. | |
