MVPF: Multitask Visual Perception Framework for Robust Sewer Defect Segmentation Under Challenging Environmental Conditions

Regular assessment of underground sewer pipe is vital for the maintenance of the sewer system. Previous studies mainly focus on automated defect detection in closed-circuit television images based on computer vision techniques. Nevertheless, challenging environmental conditions inside pipelines, such as liquid surface reflection and uneven illumination conditions, compromise the accuracy of automated sewer assessment systems. In order to address these issues, a multi-task visual perception framework (MVPF) is proposed for robust sewer defect segmentation. Firstly, we employ Pipe-SegNet to perform pipe defect segmentation task. This architecture, comprising dual-encoder branches, demonstrates strong robustness against diverse defect types inside pipelines. Subsequently, in order to cope with challenging environmental conditions, water segmentation task and dominant vanishing point detection task are assigned to Pipe-WSNet and Pipe-DVPNet, respectively. Liquid-air interface M<sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">w</sub> and weak illumination region M<sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">d</sub> can be derived from these two tasks. Eventually, M<sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">w</sub> and M<sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">d</sub> are further integrated to guide the training of Pipe-SegNet and refine its inference results, adhering to the distribution patterns of real-world defects. Extensive experiments based on practical engineering demonstrate that Pipe-SegNet achieves a more superior segmentation performance across multiple defect categories compared to other counterparts, and the mIoU of segmentation results is significantly improved by 4.07% through the integration of Pipe-WSNet and Pipe-DVPNet.