A Study on Deep Learning-Based Low-Light Image Enhancement

국문 초록 (Abstract)

본 논문은 저조도 영상에서 어두운 영역을 복원하고 향상 결과의 안정성을 유지하는 문제를 중심으로 연구를 진행하였다. 저조도 영상 향상은 단순히 영상을 밝게 만드는 작업이 아니다. 어두운 영역의 가시성을 높이면서도 색상, 구조적 세부 정보, 노이즈, 밝은 영역의 안정성을 함께 고려해야 하는 복원 문제에 가깝다. 실제 저조도 영상에서는 그림자 정보 손실, 불균일한 조명, 노이즈 증가, 밝은 영역 주변의 불안정성이 동시에 나타나는 경우가 많기 때문이다.
본 논문은 대표적인 저조도 영상 향상 방법을 검토하고, Zero-DCE, RUAS, Retinexforme를 중심으로 비교 분석을 수행하였다. 비교 결과, RUAS는 경량화 설계를 잘 보여주는 방법이고, Retinexformer는 강한 복원 성능을 보이는 방법으로 확인되었다. 반면 Zero-DCE는 상대적으로 안정적인 결과를 제공하지만, 매우 어두운 영역에서는 향상이 충분하지 않은 경향을 보였다. 이러한 관찰은 본 연구에서 Zero-DCE를 개선 대상으로 선택한 주요 이유가 되었다.
이러한 한계를 보완하기 위해 본 연구는 기존 Zero-DCE의 손실 함수 구조에 어두운 영역을 고려한 제약을 추가한 개선 방법을 제안하였다. 제안 방법은 휘도 정보를 이용하여 입력 영상의 어두운 영역을 구분하고, 임계값 기반 가중치 전략을 통해 해당 영역에 더 강한 향상 지도를 부여한다. 동시에 이미 충분히 밝은 영역에서는 불필요한 변형을 줄이도록 하여 전체 영상의 균형을 유지하고자 하였다. 이 과정에서 원래 Zero-DCE의 네트워크 구조는 변경하지 않았으며, 개선은 손실 함수 수준에서 이루어졌다.
실험은 공식 Zero-DCE 테스트 소스의 실제 저조도 이미지 79장과 LOL-v1 페어링 벤치마크를 대상으로 수행되었다. 실험 결과, 제안된 방법은 재현된 Zero-DCE 기준 모델 대비 LOL-v1에서 PSNR을 12.17에서 14.60으로, SSIM을 0.50에서 0.56으로 향상시켰으며, 어두운 영역의 평균 밝기를 28.53% 높였다. 또한 Retinexformer는 LOL-v1 및 무참조 평가에서 가장 우수한 종합 성능을 보인 반면, Zero-DCE는 조명 순서 보존 성능에서 상대적으로 더 안정적인 특성을 보였다.
전체적으로 제안 방법은 Zero-DCE 기준 모델의 안정적인 특성을 유지하면서 어두운 영역의 복원 성능을 높이는 데 효과적이었다. 본 논문의 핵심은 새로운 네트워크를 설계하는 것이 아니라, 기존 Zero-DCE 구조 안에서 어두운 영역을 더 적극적으로 반영하는 손실 함수를 구성한 데 있다. 따라서 본 연구는 대표적인 저조도 향상 방법들의 차이를 정리하고, 무참조 조건에서 어두운 영역 복원을 개선하기 위한 실용적인 방법을 제시한다.
주요어: 저조도 영상 복원, 어두운 영역 향상, 개선된 Zero-DCE, 무참조 학습, Retinexformer

다국어 초록 (Multilingual Abstract)

This thesis focuses on low-light image enhancement, especially the recovery of dark regions and the stability of enhanced results. In this study, enhancement is considered as a balance between visibility improvement and the preservation of color, structure, and visual naturalness. Real low-light images often includeshadow loss, uneven lighting, visible noise, and unstable areas near highlights. For this reason, this thesis treats low-lightenhancement as an image restoration problem rather than a simple brightness adjustment. However, existing methods, especially under zero-reference settings, may still show insufficient enhancement in genuinely dark regions.
This thesis reviews representative low-light image enhancement models and presents a comparative analysis of three methods: Zero-DCE, RUAS, and Retinexformer. The comparisonshows different model tendencies: RUAS reflects lightweight design, Retinexformer gives stronger restoration, and Zero-DCE produces more stable results. The reproduced Zero-DCE baseline keeps stable visual results, but its recovery of very dark regions is limited. This limitation motivates the proposed improvement.
To address this issue, this study proposes an improved Zero-DCE model by introducing a dark-region-aware constraint into the original loss-function framework. By using a luminance map and a threshold-based weighting strategy, the proposed method identifies insufficiently illuminated regions and strengthens enhancement in dark regions while reducing unnecessary distortion in already bright areas.
The experiments use two types of data: 79 real low-light images from the official Zero-DCE test source and the paired LOL-v1 benchmark. Quantitative results show that the proposed method improves the reproduced Zero-DCE baseline from 12.17 to 14.60 in PSNR and from 0.50 to 0.56 in SSIM, while increasing the mean brightness of dark regions by 28.53%. Among the compared methods, Retinexformer gives the strongest overall performance on LOL-v1 and in no-reference evaluation, while Zero-DCE preserves the lightness order more effectively. Additional parameter analysis further indicates that the proposed method improves dark-region visibility while keeping highlight distortion within an acceptable range.
Overall, the proposed method improves the Zero-DCE baseline while maintaining image consistency. More specifically, the contribution of this thesis lies not in redesigning the network architecture at the inference stage, but in a targeted improvement at the loss-function level. Therefore, this study not only provides a comparative reference for representative low-light image enhancement methods, but also offers a practical strategy for improving zero-reference enhancement in extremely dark regions.
Keywords: low-light enhancement, dark-region constraint, Zero-DCE improvement, RUAS, Retinexformer

목차 (Table of Contents)

• List of Tables ................................................................................................................. vii
• List of Figure .................................................................................................................. ix
• I. Introduction ..................................................................................................................1
• 1. Background and Significance of the Research ..................................................... 1
• 2. Research Status and Problem Statement .............................................................. 2
• 3. Limitations of Existing Methods .......................................................................... 3
• 4. Research Questions and Objectives ...................................................................... 4
• 5. Main Contributions ............................................................................................... 5
• 6. Organization of the Thesis .................................................................................... 6
• II. Related Theories and Research Status .....................................................................7
• 1. Overview of Low-Light Image Enhancement ...................................................... 7
• 2. Traditional Low-Light Enhancement Methods .................................................... 8
• 2.1. Histogram Equalization ........................................................................................ 8
• 2.2. Gamma Correction ................................................................................................ 9
• 2.3. Retinex-Based Methods ...................................................................................... 10
• 3. Classification of Deep Learning-Based Methods ............................................... 11
• 3.1. Supervised Paired-Data Methods ........................................................................ 11
• 3.2. Unsupervised and GAN-Based Methods ............................................................ 11
• 3.3. Zero-Reference Enhancement Methods .............................................................. 12
• 3.4. Lightweight and Efficiency-Oriented Methods .................................................. 13
• 3.5. Recent Advanced Methods ................................................................................. 13
• 4. Zero-DCE ........................................................................................................... 17
• 4.1. Core Idea of Zero-DCE ....................................................................................... 17
• 4.2. Network Structure of Zero-DCE ......................................................................... 18
• 4.3. Loss Functions of Zero-DCE .............................................................................. 20
• 4.4. Strengths and Limitations of Zero-DCE ............................................................. 21
• 5. RUAS ................................................................................................................. 22
• 5.1. Core Idea of RUAS ............................................................................................. 22
• 5.2. Retinex-Inspired Formulation and Unrolling Structure ...................................... 22
• 5.3. Cooperative Prior Architecture Search ............................................................... 23
• 5.4. Loss Design and Optimization Objective ........................................................... 24
• 5.5. Strengths and Limitations of RUAS ................................................................... 24
• 6. Retinexformer ..................................................................................................... 25
• 6.1. Core Idea of Retinexformer ................................................................................ 25
• 6.2. One-Stage Retinex-Based Formulation .............................................................. 25
• 6.3. Illumination-Guided Transformer Mechanism ................................................... 26
• 6.4. Loss Design of Retinexformer ............................................................................ 27
• 6.5. Strengths and Limitations of Retinexformer ....................................................... 27
• 6.6. Relevance of Retinexformer to This Thesis ........................................................ 28
• 7. Comparative Positioning of the Three Core Models .......................................... 28
• 8. Summary ............................................................................................................. 30
• III. Reproduction Experiments and Comparative Analysis .....................................32
• 1. Overview of the Comparative Experiments ....................................................... 32
• 2. Experimental Setting .......................................................................................... 33
• 2.1. Experimental Objective ...................................................................................... 33
• 2.2. Experimental Environment ................................................................................. 34
• 2.3. Datasets, Image Sources, and Selection Criteria................................................. 35
• 2.4. Scene Categories Used in This Chapter .............................................................. 36
• 3. Evaluation Metrics .............................................................................................. 36
• 3.1. Full-Reference Metrics ....................................................................................... 37
• 3.2. No-Reference Metrics ......................................................................................... 37
• 3.3. Metric Usage in This Chapter ............................................................................. 38
• 4. Reproduction of the Zero-DCE Baseline............................................................ 39
• 4.1. Reason for Selecting Zero-DCE as the Main Baseline ....................................... 39
• 4.2. Checkpoint Inspection and Baseline Selection ................................................... 39
• 4.3. Comparison with the Official Zero-DCE Output ................................................ 41
• 4.4. Baseline Characteristics ...................................................................................... 42
• 5. Qualitative Comparison of Representative Models ............................................ 42
• 5.1. Comparison in a General Low-Light Scene ........................................................ 42
• 5.2. Comparison in a Very Dark Scene...................................................................... 44
• 5.3. Comparison in a Complex-Light Scene .............................................................. 45
• 6. Quantitative Comparison .................................................................................... 47
• 6.1. Paired Quantitative Comparison on LOL-v1 ...................................................... 47
• 6.2. No-Reference Quantitative Comparison on the 79 Real Low-Light Images ...... 48
• 7. Discussion ........................................................................................................... 50
• 7.1. Overall Comparison of the Three Models .......................................................... 50
• 7.2. Why Zero-DCE Remains the Baseline for Chapter 4 ......................................... 51
• 8. Summary ............................................................................................................. 51
• IV. Improved Zero-DCE Framework with Dark-Region-Aware Constraint .........53
• 1. Overview of the Proposed Method ..................................................................... 53
• 2. Motivation for Improving Zero-DCE ................................................................. 54
• 3. Improved Zero-DCE with Dark-Region-Aware Constraint ............................... 55
• 3.1. Framework Overview ......................................................................................... 55
• 3.2. Luminance Map Construction ............................................................................ 57
• 3.3. Dark-Region Weighting Map ............................................................................. 58
• 3.4. Dark-Region-Aware Exposure Loss ................................................................... 59
• 3.5. Final Training Objective ..................................................................................... 60
• 4. Parameter Setting and Analysis .......................................................................... 61
• 4.1. Effect of Dark-Region Threshold 𝜏 ................................................................... 62
• 4.2. Effect of Dark-Loss Weight 𝜆𝑑 ......................................................................... 63
• 4.3. Checkpoint Selection of the Improved Model .................................................... 64
• 4.4. Visual Comparison Before and After Improvement ........................................... 65
• 4.5. Dark-Region Brightness Improvement Analysis ................................................ 67
• 5. Discussion of the Proposed Method ................................................................... 68
• 6. Summary ............................................................................................................. 69
• V. Experimental Results and Analysis .........................................................................71
• 1. Overview of the Experimental Results ............................................................... 71
• 2. Experimental Design .......................................................................................... 72
• 2.1. Objective of the Evaluation ................................................................................ 72
• 2.2. Compared Methods ............................................................................................. 72
• 2.3. Datasets and Evaluation Metrics ......................................................................... 73
• 3. Comparison with the Baseline Zero-DCE .......................................................... 73
• 3.1. Qualitative Comparison with the Baseline ......................................................... 74
• 3.2. Quantitative Comparison with the Baseline ....................................................... 77
• 3.3. Dark-Region Mean Brightness Improvement Analysis ...................................... 78
• 4. Comparison with Representative Methods ......................................................... 79
• 4.1. Paired Evaluation Results ................................................................................... 80
• 4.2. No-reference Evaluation Results ........................................................................ 81
• 5. Inference-Time Comparison ............................................................................... 82
• 6. Scene-wise Analysis ........................................................................................... 83
• 6.1. Analysis in General Low-light Scenes ................................................................ 84
• 6.2. Analysis in Very Dark Scenes ............................................................................ 85
• 6.3. Analysis in Complex-light Scenes ...................................................................... 87
• 7. Ablation Study on the Effect of 𝐿𝑑𝑎𝑟𝑘 ............................................................. 87
• 8. Supplementary Analysis on Highlight-related Limitation.................................. 88
• 9. Summary ............................................................................................................. 92
• VI. Conclusion and Future Work ................................................................................94
• 1. Overall Summary of the Study ........................................................................... 94
• 2. Main Findings and Contributions ....................................................................... 95
• 2.1. Answers to the Research Questions .................................................................... 95
• 2.2. Comparative Findings ......................................................................................... 96
• 2.3. Methodological Contribution .............................................................................. 96
• 2.4. Experimental Contribution ................................................................................. 97
• 3. Limitations of the Present Study ........................................................................ 98
• 4. Future Work ........................................................................................................ 99
• 4.1. Highlight-Preserving Enhancement for Complex-Light Scenes ......................... 99
• 4.2. Joint Optimization of Dark and Bright Regions ................................................. 99
• 4.3. Broader Evaluation on More Diverse Real Scenes ............................................. 99
• 4.4. Real-Time Deployment and Practical Validation ............................................. 100
• 4.5. Further Improvement of Lightweight Baseline Models .................................... 100
• 5. Final Remarks ................................................................................................... 100
• References ....................................................................................................................102
• 초록 ...............................................................................................................................105
• Abstract ........................................................................................................................108
• Acknowledgements ......................................................................................................110