Low-light image enhancement method based on retinex theory and dual-tree complex wavelet transform
Abstract Image quality in low-light environments is typically poor, characterized by low contrast, reduced visibility, and sensor noise. These issues not only impair human visual perception but also pose significant challenges for computer vision tasks. To address these problems, various Low-light I...
Saved in:
| Main Authors: | , , , , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Springer
2025-06-01
|
| Series: | Journal of King Saud University: Computer and Information Sciences |
| Subjects: | |
| Online Access: | https://doi.org/10.1007/s44443-025-00102-6 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Abstract Image quality in low-light environments is typically poor, characterized by low contrast, reduced visibility, and sensor noise. These issues not only impair human visual perception but also pose significant challenges for computer vision tasks. To address these problems, various Low-light Image Enhancement (LIE) algorithms have been developed in recent years, including histogram-based methods, gamma correction, and approaches based on Retinex theory. Although supervised learning methods have made significant progress in the LIE field, their reliance on paired low-light and normal-light images for training poses practical challenges. Unsupervised methods, while reducing dependency on paired images, often suffer from limited data priors, leading to artifacts and subpar image quality. Furthermore, existing Convolutional Neural Network (CNN)-based methods are limited in capturing long-range dependencies and non-local self-similarities. Although Vision Transformers (ViTs) excel at modeling non-local information, they inadequately perceive frequency information in low-light images, which is crucial for image enhancement. Therefore, this paper proposes a novel LIE framework based on Retinex theory and Dual-Tree Complex Wavelet Transform (DTCWT). This method introduces multi-scale spectral layers for effective frequency decomposition and designs a spectral optimization network that calibrates low-frequency and high-frequency information through tensor fusion and Einstein hybrid methods. Additionally, we design reflectance consistency loss and Retinex decomposition loss to enhance the model’s generalization capability under complex lighting conditions. Extensive experiments on the MIT Adobe FiveK and SICE datasets demonstrate that the proposed unsupervised method outperforms existing state-of-the-art techniques in terms of brightness enhancement, color fidelity, noise suppression, and contrast improvement, showcasing its superior performance and potential in low-light image enhancement. |
|---|---|
| ISSN: | 1319-1578 2213-1248 |