Controlled Benchmarking and Component-Aware Ablation for Railway Viaduct Structural and Damage Segmentation

SAFENET: Image-Based Infrastructure Classification of Bridges and Viaducts for Seismic Risk Assessment

Transportation networks are critical for emergency response after earthquakes, but national-scale bridge and viaducts inventories often lack vulnerability-related attributes such as material and structural system. This paper presents an image-based approach, developed within the SAFENET project, to automatically classify bridges/viaducts according to a practical Material-Structure (MS) labeling scheme that reduces sparsity compared to finer taxonomies that also include construction period. Using a Portuguese bridge image dataset, we compare three visual model families for Material-Structure classification: a ResNet-50 convolutional baseline, a self-supervised vision transformer (DINOv2-Large), and a contrastive vision encoder (CLIP). Models are evaluated with a strict 5-by-5 Nested CrossValidation (NCV) protocol with bridge-level splits to prevent information leakage across train and test sets. Results show that DINOv2 achieves the best overall performance, with a mean accuracy of 0.903, a macro-F1 of 0.773, and a weighted-F1 of 0.897, outperforming ResNet-50 and CLIP especially on minority classes. These findings support the use of self-supervised vision transformers to enrich bridge inventories from imagery and to provide scalable inputs for regional seismic risk assessment.

Intelligent damage monitoring system for hydraulic structures with the detection–segmentation collaborative framework

Semantic segmentation and quantitative analysis of tunnel cracks and water leakage using a TransUNet framework

As a vital component of structural health monitoring, the detection of cracks and water leakage in tunnel linings is essential for ensuring structural durability and operational safety. However, due to complex site conditions, such as non-uniform illumination, surface texture interference, and the slender, blurred nature of defects, traditional manual inspections and threshold-based algorithms often fail to provide reliable damage identification. To address these challenges, this study proposes an end-to-end semantic segmentation framework based on TransUNet. By integrating the local feature extraction of convolutional neural networks (CNNs) with the global dependency modeling of Transformers, the framework significantly enhances the characterization of multi-scale defects and boundary features. A comprehensive dataset comprising public benchmarks and real-world engineering images was developed using a standardized preprocessing and validation pipeline. The proposed method was systematically evaluated against state-of-the-art models like U-Net and DeepLabv3 + . Experimental results demonstrate that the TransUNet framework achieves an IoU of 71.57% for crack segmentation and a Precision of 91.51% for water leakage. Crucially for engineering applications, the geometric error for length and area measurements is maintained within 5%, while the inference latency remains under 200 ms. In terms of precision, boundary preservation, and geometric consistency, the proposed method shows clear advantages over the comparison models, while U‑Net exhibits stronger region overlap for water leakage detection. Overall, the method meets the requirements of offline inspection and near-real-time applications. This data-driven approach provides a robust technical foundation for tunnel defect detection and subsequent maintenance decision-making.

MDeepLab: a real-time lightweight network for accurate road crack segmentation

Experimental results demonstrate that, compared with mainstream semantic segmentation models, MDeepLab significantly reduces the number of parameters while maintaining high segmentation accuracy, exhibiting promising engineering application value for automated road crack detection.

A Novel Lightweight Framework for Real-Time Pavement Crack Segmentation Based on Knowledge Distillation

Accurate pavement crack segmentation is essential for structural health monitoring, yet existing methods often face a trade-off between segmentation accuracy and computational efficiency. To address this issue, a novel teacher–student framework, termed RTCS-T and RTCS-S, is proposed. The teacher network RTCS-T is constructed based on the Swin Transformer to capture long-range dependencies and multi-scale contextual information. To further enhance crack representation, a strip refinement module is introduced to model directional structural features, while a cascaded atrous spatial pyramid pooling module is employed to improve multi-scale feature aggregation. Based on the teacher network, a lightweight student model RTCS-S is developed by using depthwise separable convolutions to achieve efficient inference. In addition, a foreground-aware and boundary-aware knowledge distillation strategy is introduced to guide the transfer of structural and contextual information from the teacher to the student. Experiments on the Crack500, DeepCrack, and CFD datasets demonstrated competitive performance against representative segmentation models. On CFD, RTCS-S achieved an F1 Score of 0.7514 and an mIoU of 0.7962. Notably, RTCS-S required only 1.82 M parameters and 1.13 GFLOPs and achieved a model inference speed of 680 FPS on an RTX 4090 GPU. When deployed on an RDK X5 edge-computing platform, the complete pipeline achieved an end-to-end throughput of 34 FPS, with an average latency of approximately 29.4 ms and peak memory consumption of 1.8 GB. These results demonstrate that the proposed framework provides an efficient solution for automated pavement crack detection and shows strong potential for practical road inspection applications.

Automated Visual Inspection of Bridge Defect Segmentation Using Large-Scale Pretrained Models

Routine visual inspections of bridges are safety-critical activities that are still manual, time-consuming, and subjective to the inspector’s interpretation. This paper presents a scalable automated visual inspection pipeline for semantic segmentation of bridge defects, targeting the 19-class dacl10k benchmark dataset. We perform the first systematic comparison of pretraining paradigms on dacl10k, including CNN -supervised, ViT-supervised, masked image modelling, and self-supervised learning across 10 configurations under a controlled training protocol. The pretraining paradigm consistently dominates architectural choice, with DINOv2-L, pretrained on 142 million unlabelled images, achieving a mean Intersection-over-Union (mIoU) of 49.16%. Applying our native multi-label training approach to EVA-02-L, the dacl10k challenge-winning backbone, achieves 48.97% mIoU versus their 47.80% single-model result, demonstrating that training design is an independent performance factor. A three-model ensemble achieves 51.08% mIoU, exceeding the challenge winner’s score by using half the number of models. A prototype inspection system (InSpectralytiX) is deployed in a HuggingFace Gradio Space, demonstrating end-to-end feasibility from raw image to a per-class defect map. The future work targets automated condition scoring for bridge asset management integration, supporting structural health monitoring at the local level when performed repeatedly.