Comparative Evaluation of Image Segmentation Models for Bridge Components

다국어 초록 (Multilingual Abstract)

Accurate alignment between Building Information Modeling (BIM) models and real bridge
scenes is essential for effective augmented reality (AR) applications in inspection, maintenance, and health monitoring. Yet, outdoor bridge environments present severe challenges for
vision-based tracking, including complex backgrounds, variable illumination, occlusions, and
geometric variability of components. Conventional AR tracking approaches, which rely on sparse feature matching, markers, or coarse geometry, often fail to provide stable and precise registration at the component level, limiting the reliability of BIM overlays for engineering
decision-making.
Deep learning-based image segmentation has the potential to address this limitation by
providing dense, component-level masks that can serve as strong geometric and semantic
constraints for pose estimation. However, existing research typically evaluates segmentation models only on static accuracy metrics and often focuses on damage or generic object classes
rather than bridge components in BIM AR workflows. There is a lack of systematic analysis of
how different segmentation architectures CNN-based semantic segmentation, transformer-
based models, and real-time single-stage detection segmentation networks trade off
segmentation accuracy, inference speed, and mask quality when used as the front end for
BIM-based AR tracking and pose estimation in real bridge inspection scenarios.

국문 초록 (Abstract)

교량 구성 요소의 정확한 분할은 토목 구조물, 특히 시각화, 정렬 및 자동 검사와 같은 작업에서 증강 현실(AR) 애플리케이션의 핵심 요소입니다. GPS 또는 기준 마커와 같은 기존 AR 추적 방식은 특히 복잡한 교량 환경에서 정밀도와 안정성이 부족한 경우가 많습니다. 딥러닝 기반의 시맨틱 분할은 구조 요소를 감지하고 분리하는 비전 기반 대안을 제공하며, 이는 AR 시스템에서 이러한 구성 요소의 안정적인 자세 추정을 위한 필수적인 첫 단계입니다.주제어(키워드).
본 논문은 교량 구성 요소의 의미론적 분할을 위한 세 가지 딥러닝 모델, 즉 DeepLabV3+ (ResNet-50 및 ResNet-101 백본 사용), YOLOv11n-Seg, 그리고 트랜스포머 기반 SegFormer (MIT-B0~B3)에 대한 비교 평가를 제시합니다. 이러한 모델들은 분할 정확도, 경계 품질, 파라미터 크기, 그리고 후속적인 자세 추정 및 AR 배포에 중요한 실시간 성능 측면에서 분석되었습니다. 결과는 DeepLabV3+가 가장 높은 분할 정확도를 달성하는 반면, YOLOv11n-Seg는 현장 사용에 적합한 부드럽고 명확한 경계를 제공하며 가벼운 배포가 가능하다는 것을 보여줍니다. SegFormer 모델은 특히 작거나 덜 두드러진 특징을 분할하는 데 있어 뛰어난 일반화 성능을 나타냅니다. 본 연구는 AR 기반 교량 모니터링 워크플로에서 최적의 분할 백본을 선택하는 데 유용한 정보를 제공합니다.

목차 (Table of Contents)

• List of Figures iv
• List of Tables vi
• Abstract vii
• Chapter 1. Introduction 1
• Chapter 2. Literature Review 3
• 2.1. Introduction to Segmentation Model in Construction Industry 3
• 2.1.1. Semantic Segmentation for Components Level Understanding 3
• 2.2. CNN-Based Segmentation Models 4
• 2.2.1. U-Net Encoder-Decoder Archtecure with Skip Connection 5
• 2.2.2. Mask R-CNN Two Stage Instant Segmentation with High Precision 6
• 2.2.3. PSPNet Pyramid Scene Parasing for Global Context 8
• 2.2.4. Feature Pyramid Scene Network(FPN) 9
• 2.2.5. DeeplabV Series Multi Scale Context Through ASPP 11
• 2.2.1. YOLO Series Model Real Time Detection and Instant Segmentation 14
• 2.8. Transformer Based Segmentation Model 17
• 2.8.1. Mask2Former Universal Archicture for Panoptic Segmentation 17
• 2.8.2. Segment Anything Model (SAM) 19
• 2.8.3. Why SAM Fails for Real Time AR Application 20
• 2.8.4. SegFormer model Architecture 20
• 2.9. AR in Construction 23
• Chapter 3. Research Problem and Objectives 24
• Chapter 4. Methodology 26
• 4.1. Dataset Preparation 26
• 4.1.1. Dataset Variant and Dataset Split proportions for Training 28
• 4.1.2. Dataset Annotation 28
• 4.2. Overview of Selected Segmentation Models 29
• 4.3. Hyperparameter Configuration 31
• 4.4. Model Backbone and Key Feature 33
• 4.5. Evaluation Metrics 34
• Chapter 5. Experiments and Results 20
• 5.1. Training Results and Analysis 35
• 5.1.1. Training Results of YOLO11n-Seg on Dataset A 35
• 5.1.2. Training Results of YOLO11n-Seg on Dataset B 37
• 5.1.3. Training Results of DeeplabV3+ ResNet on Dataset A 39
• 5.1.4. Training Results of DeeplabV3+ ResNet on Dataset B 41
• 5.1.5. Training Results of SegFormer-B3 on Dataset A 44
• 5.1.5. Training Results of SegFormer-B3 on Dataset B 47
• 5.2. Quantitative Results Comparison 50
• 5.2.1. Quantitative Results Performance Comparison on Dataset A 50
• 5.2.2. Quantitative Results Performance Comparison on Dataset B 52
• 5.2.3. Per-Class Evaluation Metrics Performance on Dataset A 54
• 5.2.4. Per-Class Evaluation Metrics Performance on Dataset B 55
• 5.3. Qualitative Results Comparison among Three Algorithms 57
• 5.3.1. Qualitative Results Comparison on Dataset A 57
• 5.3.2. Qualitative Results Comparison on Dataset B 59
• 5.4. Application of YOLO11n-Seg for Pose-estimation 62
• 5.5. Discussion 63
• 5.6. Hardware Specification 65
• Chapter 6. Conclusion 66
• References 68
• Korean Abstract 77