CH:ALK - Rapid automatic labeling toolkit to develop training images for concrete damage segmentation models

Abstract

With the growing demand for automated structural inspection due to the aging of civil infrastructure, deep segmentation models have been increasingly adopted with the imaging of structures. However, training the models using common supervised learning requires labeled data, and traditional manual labeling is labor-intensive, inconsistent, and time-consuming. This study introduces CH:ALK (Concrete Highlighter: Accelerated Labeling Toolkit), a rapid labeling toolkit designed to produce fast, accurate, and consistent training images for supervised learning of damage segmentation models. CH:ALK integrates automatic labeling (AL) using pre-trained CGNet (Context-Guided Network) and SAM (Segment Anything Model) to label four types of concrete damage: cracks, efflorescence, rebar exposure, and spalling. CH:ALK supports pixel-level AL that can be followed by manual correction via brush tools in an intuitive GUI. Performance validation using 80 images labeled by four users demonstrated an average time reduction of 87.97 %, accuracy of 67.07 % (mIoU), and inter-user consistency of 78.44 %, compared with traditional manual labeling (ML). Furthermore, two segmentation models, CGNet and DeepLabV3+, trained with AL data showed comparable performance to those trained with ML data. CH:ALK offers a scalable solution for developing high-quality labeled datasets for civil infrastructure inspection.