Automated multi-category tunnel damage detection and report generation from ultra-high-resolution panoramic laser images

challenge due to the inherent subjectivity and time demands of manual inspections. Although reality capture technology allows for digital representation of as-is condition of assets, converting these rich data sources into actionable risk assessments demands still requires innovative solutions. In this paper, we introduce a comprehensive, web-based automated framework that uses ultra-high-resolution (UHR) panoramic tunnel images to automatically generate detailed damage records and risk assessment reports. A significant challenge in this domain is the observation that damage regions often lack sharply defined boundaries; instead, they exhibit gradual, blurred transitions, which is not well-suited to conventional segmentation evaluation. To address this, we formally define the challenge of inconsistency of damage annotation in complex real-world scenarios and propose a novel evaluation metric: Intersection over Union with buffer zone (IoUb). This metric relaxes the rigid boundary precision requirements of traditional evaluation methods, focusing more on capturing the overall damage. We evaluated several instance segmentation algorithms and recommend adopting a lower confidence threshold, as it reduces missed detections without significantly increasing false positives. We introduce post-processing methods that aggregate the predictions from multiple inferences to meet the demands of processing UHR panoramic images, resulting in a 3% improvement in Macro IoU and IoUb, along with a 90% damage recall. Experimental results on Italian road tunnels demonstrate that our framework enhances automated damage detection. We then categorize damage severity using a statistically grounded methodology, enable natural language queries of statistical damage results, and handle visualization and report export, all within a single end-to-end web-based platform. The proposed framework significantly enhances the efficiency of professionals in planning and monitoring ageing tunnel assets. Our code is available at https://github.com/zxy239/Auto-damage-report-generation