본문 바로가기 주메뉴 바로가기
국회도서관 홈으로 정보검색 소장정보 검색

결과 내 검색

동의어 포함

목차보기

Title Page 2

Contents 5

국문요약 9

Chapter 1. Preliminary Research: MSDGAN-Multi-Scale Dilated Generative Adversarial Network for Smoke Removal and Image Restoration 11

1.1. Introduction 11

1.2. Related Work 13

1.2.1. Generative Adversarial Networks (U-Net GAN) 13

1.2.2. Residual Block Architecture 15

1.2.3. Image Inpainting Techniques 16

1.2.4. Single Image Dehazing 17

1.3. Method 18

1.3.1. Dataset 18

1.3.2. Proposed Model 20

1.3.3. Loss Function 25

1.3.4. Network Summary 30

1.4. Result 31

1.4.1. Evaluation Metrics 31

1.4.2. Baseline Comparisons 33

1.4.3. Quantitative Results 34

1.4.4. Qualitative Results 36

1.4.5. Cross-Cultural Generalization 39

1.4.6. Discriminator Local Feedback Analysis 41

1.5. Conclusion 43

Chapter 2. The Main Study: BADNet - Balanced Attention Dehazing Network for Single Image Dehazing 45

2.1. Introduction 45

2.1.1. Motivation and Background 45

2.1.2. Deep Learning Approaches and Challenges 47

2.1.3. Proposed Method Overview 54

2.2. Related Research 55

2.2.1. Traditional Image Dehazing Methods 56

2.2.2. Learning-Based Dehazing Methods 60

2.2.3. Loss Function Design and Optimization Strategies 70

2.2.4. Datasets and Benchmarking in Dehazing 71

2.2.5. Applications and Real-World Deployments of Dehazing 73

2.3. Proposed Method 75

2.3.1. Overall Model Description 75

2.3.2. Core Building Blocks: DMSB and BCPAB Overview 78

2.3.3. Dilated Multi-scale Block (DMSB) 80

2.3.4. Balanced Channel-Pixel Attention Block (BCPAB) 84

2.3.5. Balanced Attention Loss (BAL) 88

2.3.6. Network Architecture and Training Strategy 90

2.4. Experimental 92

2.4.1. Experimental Setup 92

2.4.2. Quantitative Results 96

2.4.3. Qualitative Comparison 97

2.4.4. Ablation Study Results 102

2.4.5. Computational Analysis and Limitations 105

2.5. Conclusion 109

Reference 112

ABSTRACT 126

List of Tables 7

Table 1.1. Comparison with other Models 34

Table 2.1. Comparison of dehazing methods on ITS, OTS and RESIDE6K 96

Table 2.2. Ablation Study Results 103

List of Figures 8

Figure 1.1. Smoke-Free, Smoke Image and Smoke Mask For each Density 20

Figure 1.2. Generator model architecture 21

Figure 1.3. MSDRB module architecture 22

Figure 1.4. Proposed Unet-GAN discriminator 24

Figure 1.5. Proposed Network Summary 30

Figure 1.6. Results of each model for Landmark images 36

Figure 1.7. Discriminator Local Output with Landmark 41

Figure 2.1. Impact of Weather Conditions on Traffic Fatalities 2025 45

Figure 2.2. Foggy Airport Scene 46

Figure 2.3. Various Approaches for Dehazing 47

Figure 2.4. Atmospheric Scattering Model Based on Koschmieder's Law 56

Figure 2.5. Various Physics-Based Model 57

Figure 2.6. Uniform, Non-Uniform Haze Scene 59

Figure 2.7. CNN-Based Approaches for Dehazing 60

Figure 2.8. MSDGA: CNN-based Single Image Dehazing Model 61

Figure 2.9. The proposed BADNet architecture with DMSB and Balanced... 75

Figure 2.10. Architecture overview of the two core building blocks. Left:... 78

Figure 2.11. Detailed architecture of the Dilated Multi-scale Block (DMSB). The... 80

Figure 2.12. Detailed architecture of the Balanced Channel-Pixel Attention... 84

Figure 2.13. Comparison Qualitative Results 98

Figure 2.14. Qualitative Analysis of Road-Visible Outdoor Environment 98

Figure 2.15. Comparison patches showing texture details 99

Figure 2.16. Comparison focusing on color accuracy in city street scenes 100

Figure 2.17. Comparison focusing on color accuracy in road scenes 100

Figure 2.18. Comparison Indoor Hazed Environments 101

초록보기

 최근 이미지 디헤이징(dehazing)은 자율 주행, 감시 시스템, 의료 영상 등 다양한 컴퓨터 비전 응용에서 중요한 연구 분야로 자리 잡고 있습니다. 기존 기법들은 물리적 모델 기반 접근법이나 딥러닝을 활용한 방법론을 적용하였으나, 단일 이미지에서 다중 스케일과 픽셀 수준의 정보를 효과적으로 활용하는 데 한계가 있었습니다.

본 연구에서는 BADNet (Balanced Attention Dehazing Network)을 제안하며, 강건한 단일 이미지 디헤이징을 위해 균형 주의 메커니즘(Balanced Attention Mechanisms)을 도입하였습니다. 먼저, 챕터 1에서는 로컬 및 글로벌 discriminator를 활용하여 Dilated Multi-Scale Block (DMSB)의 효과를 연구하였고, 이를 기반으로 확장된 디헤이징 기법을 개발하였습니다. 이후 연구인 챕터 2에서는 다중 스케일 정보와 픽셀 수준 최적화를 활용한 모델인 BADNet을 제안하였으며, 주요 구성 요소로 (1) 균형 채널-픽셀 주의 블록(BCPAB, Balanced Channel-Pixel Attention Block)을 적용하여 정보의 집중적 처리, (2) 균형 주의 손실(BAL, Balanced Attention Loss)을 통해 학습을 최적화하였습니다.

실험 결과, 본 연구에서 제안한 모델은 기존 최첨단(SOTA) 디헤이징 기법 대비 우수한 PSNR 및 SSIM 성능을 기록하였으며, 다양한 환경에서도 안정적인 성능을 유지함을 입증하였습니다. 본 연구는 이미지 복원 기술의 새로운 가능성을 제시하며 실세계 응용에서도 높은 실용성을 갖추고 있습니다