In sewer pipe Closed-Circuit Television (CCTV) inspection, accurate defect segmentation is fundamental for effective condition assessment, yet remains challenging due to defects’ heterogeneous appearances and subtle boundary characteristics. Current deep learning methods often compromise between model interpretability and feature representation capability when processing such complex visual patterns. To address these limitations, we propose PipeUKAN, a novel semantic segmentation framework based on the U-KAN architecture, which introduces: (1) Kolmogorov-Arnold Net-works (KANs) to enhance nonlinear feature learning while preserving model transparency, (2) an Atrous Spatial Pyramid Pooling (ASPP) module for multiscale defect context capture, and (3) a lightweight Edge Attention Module (EAM) for high-precision boundary refinement. We evaluate PipeUKAN on CCTV inspection images featuring four common pipeline defects: racks, hidden lateral connections, misalignments, and leaks. Extensive experiments demonstrate that our method outperforms the original U-KAN across multiple segmentation metrics, achieving superior accuracy and robustness.