표제지
국문초록
목차
1. 서론 14
1.1. 연구 배경 14
1.2. 연구 동향 16
1.3. 연구 목적 및 범위 21
2. 이론적 배경 23
2.1. 딥러닝 모델의 학습 23
2.2. 딥러닝 모델 학습을 위한 이미지 라벨링 방법 27
2.3. 반자동 라벨링을 위한 딥러닝 모델 36
2.3.1. CGNet(Context-Guided Network) 36
2.3.2. SAM(Segment Anything Model) 38
3. 신속 라벨링 시스템 개발 43
3.1. 신속 라벨링 시스템 개요 43
3.2. 신속 라벨링 시스템 소프트웨어 개발 44
3.2.1. 신속 라벨링 시스템 기반 이미지 라벨링 44
3.2.2. 반자동 이미지 라벨링을 위한 딥러닝 모델 학습 56
3.3. 신속 라벨링 시스템 GUI 61
4. 신속 라벨링 시스템 성능 검증 64
4.1. 신속 라벨링 성능 검증 64
4.1.1. 반자동-수동 라벨링 비교 실험 64
4.1.2. 라벨링 성능 검증 결과 67
4.2. 라벨링 방식에 따른 학습 성능 검증 81
4.2.1. 반자동-수동 라벨링 데이터에 따른 학습 비교 실험 81
4.2.2. 학습 성능 검증 결과 84
5. 결론 및 추후 연구 96
5.1. 결론 96
5.2. 추후 연구 98
참고문헌 99
Abstract 105
Table 1.1. Image labeling tool for semantic segmentation 20
Table 3.1. CGNet model training configuration 57
Table 3.2. CGNet model training data set each damage 57
Table 3.3. CGNet model test results each damages 59
Table 4.1. Crack labeling comparative experiment: speed verification 69
Table 4.2. Efflorescence labeling comparative experiment: speed verification 69
Table 4.3. Rebar-Exposure labeling comparative experiment: speed verification 70
Table 4.4. Spalling labeling comparative experiment: speed verification 70
Table 4.5. Crack labeling comparative experiment: accuracy verification 74
Table 4.6. Efflorescence labeling comparative experiment: accuracy verification 74
Table 4.7. Rebar-Exposure labeling comparative experiment: accuracy verification 75
Table 4.8. Spalling labeling comparative experiment: accuracy verification 75
Table 4.9. Crack labeling comparative experiment: consistency verification 78
Table 4.10. Efflorescence labeling comparative experiment: consistency verification 78
Table 4.11. Rebar-Exposure labeling comparative experiment: consistency verification 79
Table 4.12. Spalling labeling comparative experiment: consistency verification 79
Table 4.13. Train data 83
Table 4.14. Test data 83
Table 4.15. Training configuration for each model 83
Table 4.16. Evaluation Result: Crack (SA: Semi-Automatic data, M: Manual data) 85
Table 4.17. Evaluation Result: Efflorescence (SA: Semi-Automatic data, M: Manual data) 88
Table 4.18. Evaluation Result: Rebar-Exposure (SA: Semi-Automatic data, M: Manual data) 91
Table 4.19. Evaluation Result: Spalling (SA: Semi-Automatic data, M: Manual data) 94
Figure 2.1. Single layer perceptron 25
Figure 2.2. Multi layer perceptron 26
Figure 2.3. Fully connected network for MNIST classification 28
Figure 2.4. Convolutional Neural Network Architecture 28
Figure 2.5. Convolutional Layer by Filter(Kernel) 30
Figure 2.6. Rectified Linear Unit (ReLU) 30
Figure 2.7. Output for each model (a) Image classification, (b) Object Detection, (c) Semantic Segmentation 32
Figure 2.8. Image Classification ImageNet Dataset 33
Figure 2.9. Object Detection Pascal Dataset 33
Figure 2.10. Semantic Segmentation Cityscapes Dataset: Example Münster 35
Figure 2.11. Superpixel based stuff Labeling 35
Figure 2.12. Polygon labeling 35
Figure 2.13. Context Guided Network 37
Figure 2.14. Context Guided block 37
Figure 2.15. Build a foundation model for segmentation by introducing three interconnected components 39
Figure 2.16. Segment Anything Model Overview 40
Figure 3.1. Semi-Automatic image labeling software: CGNet 46
Figure 3.2. Semi-Automatic image labeling flowchart: CGNet 47
Figure 3.3. Visualized semi-automatic labeling results after mapping (Crack) 47
Figure 3.4. Labeled data as binarization 47
Figure 3.5. Semi-Automatic image labeling software: SAM 50
Figure 3.6. Semi-Automatic image labeling flowchart: SAM 50
Figure 3.7. Re-detection of damage through prompt settings (a) Select damage object(Red Star: Foreground),... 51
Figure 3.8. Freehand brush labeling tools (a) brush tool, (b) erase tool 54
Figure 3.9. Label opacity, (a) α=0.8, (b) α=0.2 55
Figure 3.10. Filling hole labeling, (a) outline labeling, (b) filling hole 55
Figure 3.11. Intersection over Union 59
Figure 3.12. CGNet model test data detection results (a) Crack, (b) Efflorescence, (c) Rebar-Exposure, (d) Spalling 59
Figure 3.13. Loss graph (a) Crack, (b) Efflorescence, (c) Rebar-Exposure, (d) Spalling 60
Figure 3.14. Rapid labeling system (CHALK, Concrete Highlighter: Accelerated concrete damage Labeling toolKit) 62
Figure 3.15. Create project file (a) Set project name and directory, (b) Set damage class and color 63
Figure 4.1. Data set for labeling comparative experiment (a) Crack, (b) Efflorescence, (c) Rebar-Exposure, (d) Spalling 66
Figure 4.2. Overlap rate 66
Figure 4.3. Results according to labeling method (a) Semi-Automatic labeling, (b) Manual labeling 71
Figure 4.4. Semi-Automatic labeling results according to deep learning model, (a) CGNet, (b) SAM 73
Figure 4.5. Test data detection results for each model (Crack) 85
Figure 4.6. Loss graph for each model (Crack) 86
Figure 4.7. Test data detection results for each model (Efflorescence) 88
Figure 4.8. Loss graph for each model (Efflorescence) 89
Figure 4.9. Test data detection results for each model (Rebar-Exposure) 91
Figure 4.10. Loss graph for each model (Rebar-Exposure) 92
Figure 4.11. Test data detection results for each model (Spalling) 94
Figure 4.12. Loss graph for each model (Spalling) 95