Low‐light image enhancement based on Retinex decomposition and adaptive gamma correction

Abstract

Low-light images suffer from poor visibility and noise. In this paper, a low-light image enhancement method based on Retinex decomposition is proposed. A pyramid network is first utilized to extract multi-scale features to improve the quality of Retinex decomposition. Then the decomposed illumination is refined via an adaptive Gamma correction network to handle non-uniform illumination, while the decomposed reflectance is refined with a lightweight network. Finally, the enhanced image is obtained by element-wise multiplication between the refined illumination and reflectance components. Quantitative and qualitative experiments demonstrate the superiority of our method over state-of-the-art image enhancement methods. 1 INTRODUCTION Low-light images are usually degraded by low contrast and noise, which impede many computer vision tasks, such as tracking and detection. Therefore, improving the quality of lowlight images remains an essential task in computer vision. The main task of low-light images enhancement methods is to create visual pleasing and informative images, which are of high contrast and noise-free. There are many attempts to enhance low-light images, such as the histogram equalisation (HE) based methods [1-4], dehazing based methods [5, 6], Retinex decomposition based methods [7-9], and learning based methods [10-12]. Histogram equalisation(HE) [1] is one of the most simplest and widely utilized methods for image enhancement. Local histogram equalisation(LHE) [3] firstly utilize a predefined window to perform the histogram equalisation locally. Another variation [2] combines global and local hierarchical levels of HE by incorporating local features around a neighborhood for each pixel to aid equalisation. These two variation methods improve performance by sacrificing computational complexity. Jung et al. [4] present a pixel mapping function based on probability through HE, which could produce pleasing results in large smooth regions; however, it is ineffective to restore sharpness at the edges. It should be mentioned that the HE-based methods This is an open access article under the terms of the Creative Commons Attribution License, which permits use, distribution and reproduction in any medium, provided the original work is properly cited.

Unpaired strong-light image augmentation network with illumination curve

“…The current image enhancement methods are not divorced from the traditional image enhancement methods 10 . Most of them combine Retinex theory with neural network.…”

FDMLNet: A Frequency-Division and Multiscale Learning Network for Enhancing Low-Light Image

“…Hence, the LLIE problem can be viewed as the illumination component estimation. On the basis of this assumption, LR3M [ 18 ], a fast Retinex-based algorithm [ 8 ], Poisson noise aware Retinex model [ 9 ], Retinex-based variational framework [ 10 ], and other methods [ 11 , 41 ], have been reported to yield satisfying images. However, the enhanced results exhibit observable color distortion, noise enlargement, or fuzzy details.…”

γ Radiation Image Enhancement Method Based on Non-Linear Mapping

“…Illumination enhancement is mainly to improve the image brightness and contrast. As the basic step of image processing, there are many attempts to enhance undesirable illumination in scene images, such as the histogram equalization (HE) based methods [16][17][18][19], Retinex decomposition based methods [20][21][22][23], Non-linear remapping methods [24,25] and deep learning based methods [23,[26][27][28].…”

“…There are two main strategies of deep learning based illumination enhancement methods, the first one is generative adversarial networks based (GAN) method [26,[43][44][45], which generatively estimating the illumination to obtain the enhancement results, but these methods require very complete training data, and its enhancement effects are very prone to distortion. Another way to enhance image by estimating the adjustment parameters such as light level remapping curve and pixel remapping matrix, these methods [23,27,28,46,47] require less training data, but its enhancement results are still easily limited by the reliable data in the patricidal scenario. Meanwhile, the efficiency of deep learning based methods is also limited by the computing resources of the platform due to the presence of a large number of floating-point operations.…”