Maintenance-Oriented Tunnel Digital Model Generation via Panoptic Segmentation of Ultra-High-Resolution Images

Ultra-high-resolution (UHR) panoramic imaging enables detailed capture of tunnel surface conditions. Nevertheless, damage detection and reporting still relies on manual annotations after the inspection, as existing automated algorithms typically require downsampling (losing fine details) or patch-based processing (loosing globalcontext) to handle the massive computational load of UHR data. This trade-off restricts the effective use of such high-fidelity data, leaving comprehensive reporting reliant on manual annotation. To address this gap, a novel framework has been proposed that directly operates on UHR panoramic images for automated damage detection and 3D reconstruction employing Building Information Modelling to createdigital model with annotated defects. At its core is a flexible segmentation architecture with a side network that enables context extraction from larger image patches, supporting panoptic segmentation of images over 6K resolution and accurate detectionof tunnel components and five main damage types. With the Segment Anything Model 2 backbone, performance further improves, raising panoptic quality from 53.14 to 56.99. Industry Foundation Classes open data format has been expended to support the standardized and interoperable representation of damaged tunnel, based on which an as-damaged BIM model is constructed to effective record and visualize inspection outcomes and quantitative significance level, thereby enhancing the efficiency of tunnel monitoring and management.