Papers
Topics
Authors
Recent
Gemini 2.5 Flash
Gemini 2.5 Flash
119 tokens/sec
GPT-4o
56 tokens/sec
Gemini 2.5 Pro Pro
43 tokens/sec
o3 Pro
6 tokens/sec
GPT-4.1 Pro
47 tokens/sec
DeepSeek R1 via Azure Pro
28 tokens/sec
2000 character limit reached

Deep Retinex Decomposition for Low-Light Enhancement (1808.04560v1)

Published 14 Aug 2018 in cs.CV

Abstract: Retinex model is an effective tool for low-light image enhancement. It assumes that observed images can be decomposed into the reflectance and illumination. Most existing Retinex-based methods have carefully designed hand-crafted constraints and parameters for this highly ill-posed decomposition, which may be limited by model capacity when applied in various scenes. In this paper, we collect a LOw-Light dataset (LOL) containing low/normal-light image pairs and propose a deep Retinex-Net learned on this dataset, including a Decom-Net for decomposition and an Enhance-Net for illumination adjustment. In the training process for Decom-Net, there is no ground truth of decomposed reflectance and illumination. The network is learned with only key constraints including the consistent reflectance shared by paired low/normal-light images, and the smoothness of illumination. Based on the decomposition, subsequent lightness enhancement is conducted on illumination by an enhancement network called Enhance-Net, and for joint denoising there is a denoising operation on reflectance. The Retinex-Net is end-to-end trainable, so that the learned decomposition is by nature good for lightness adjustment. Extensive experiments demonstrate that our method not only achieves visually pleasing quality for low-light enhancement but also provides a good representation of image decomposition.

User Edit Pencil Streamline Icon: https://streamlinehq.com
Authors (4)
  1. Chen Wei (72 papers)
  2. Wenjing Wang (23 papers)
  3. Wenhan Yang (96 papers)
  4. Jiaying Liu (99 papers)
Citations (1,485)

Summary

Deep Retinex Decomposition for Low-Light Enhancement

The paper "Deep Retinex Decomposition for Low-Light Enhancement" presents a novel approach to low-light image enhancement via a deep learning extension of the well-established Retinex model. The authors propose a deep convolutional neural network termed Retinex-Net, which is designed to achieve superior low-light image enhancement through the decomposition of input images into reflectance and illumination. By leveraging a large scale dataset of paired low/normal-light images, this method uniquely combines image decomposition with subsequent enhancement operations, outperforming existing techniques both visually and quantitatively.

Background and Motivation

Low-light image enhancement is a crucial step in improving visibility and quality of images captured under insufficient lighting. Existing methods, including traditional histogram equalization and de-hazing techniques, provide limited success due to their reliance on handcrafted constraints and parameters. These methods often fail to generalize across varying scenes and lighting conditions. The Retinex theory offers a robust framework by decomposing images into reflectance and illumination components, but traditional Retinex-based methods also struggle due to similar limitations.

Contributions

The authors make several key contributions:

  1. Construction of LOL Dataset: The authors introduce the LOw-Light (LOL) dataset, the first large-scale collection of paired low/normal-light images captured from real scenes. This dataset is crucial for training deep learning models aimed at low-light image enhancement.
  2. Deep Learning-Based Retinex Decomposition: Retinex-Net consists of a Decom-Net for image decomposition and an Enhance-Net for illumination adjustment. This end-to-end trainable network ensures that decomposed components are optimized for subsequent enhancement operations, addressing the limitations of handcrafted models.
  3. Structure-Aware Total Variation Loss: A new loss function is introduced to ensure smoothness in the illumination map while preserving essential image structures. By weighting the total variation loss with the gradient of the reflectance map, the method retains structural details while removing texture noise.

Methodology

The Retinex-Net framework decomposes low-light images into reflectance and illumination through a Decom-Net and then enhances the illumination via an Enhance-Net. The Decom-Net leverages paired low/normal-light images during training but can operate on single low-light images during inference. The network is trained using reconstruction loss, invariable reflectance loss, and illumination smoothness loss.

For illumination smoothness, a structure-aware total variation loss is employed, which mitigates the effect of total variation in regions with strong gradients. This ensures that the resulting illumination map is both smooth and structurally consistent.

The Enhance-Net follows an encoder-decoder architecture with multi-scale concatenation to adjust illumination hierarchically, maintaining global consistency while refining local distributions. Additionally, denoising on the reflectance component is performed to counteract the amplified noise often present in low-light conditions.

Experimental Results

Extensive experiments demonstrate the effectiveness of the proposed method. The LOL dataset, coupled with additional synthetic data, provides a robust training basis for the network. The paper reports significant improvements in visual quality and image decomposition over state-of-the-art methods such as DeHz, NPE, SRIE, and LIME. Enhanced images exhibit balanced exposure, preserved details, and minimized noise without the typical artifacts seen in previous techniques.

Implications and Future Directions

The proposed method offers both practical and theoretical implications for the field of low-light image enhancement. From a practical perspective, the ability to automatically decompose and enhance images in real-world scenarios can significantly improve the performance of various computer vision applications. Theoretically, the data-driven approach to Retinex decomposition opens new avenues for exploring unsupervised learning techniques in image processing tasks.

Future work could explore further improvements in decomposition accuracy and efficiency, potential extensions to video sequences for real-time applications, and integration with other advanced deep learning frameworks to enhance performance further. Additionally, expansion of the dataset to include more diverse scenes and lighting conditions could facilitate more robust model generalization.

In summary, the paper provides a comprehensive framework for low-light image enhancement, significantly advancing the state-of-the-art through novel methodological contributions and extensive experimental validation.

Youtube Logo Streamline Icon: https://streamlinehq.com