An Improved HRNetV2-Based Algorithm for Semantic Segmentation of Corroded Regions in Urban Drainage Pipes

Urban drainage pipelines are critical infrastructure for maintaining the healthy operation of cities. Internal corrosion poses a serious threat to the structural integrity and service life of pipelines. Due to the irregular morphology, blurred edges, and frequent confusion with the background of corrosion regions, traditional image segmentation methods are limited in accuracy and robustness for such defect detection tasks. To address these challenges, this paper proposes an improved corrosion segmentation algorithm based on High-Resolution Network Version 2 (HRNetV2). HRNetV2 is employed as the backbone network, leveraging its parallel multi-branch structure to maintain high-resolution representations throughout the network and using multi-scale feature fusion to preserve fine-grained spatial details. Building upon this, we integrate a Convolutional Block Attention Module (CBAM) in the Stage 4 stage to enhance the network’s discriminative ability toward corrosion features, and embed a Lightweight Pyramid Pooling Module (LitePPM) after feature concatenation to improve the perception of corrosion patterns at different scales and strengthen contextual modeling capability. Experimental results demonstrate the superiority of the proposed method on a drainage pipe corrosion dataset, achieving 95.92% mean Intersection over Union (mIoU), 98.01% mean Pixel Accuracy (mPA), and 98.54% overall Accuracy. The final model delivers precise segmentation of corrosion regions, preserving segmentation detail without sacrificing accuracy, thus validating the effectiveness of the proposed approach.