Source-linked AI summary
Joint Enhancement and Denoising Method via Sequential Decomposition
Xutong Ren, Mading Li, Wen-Huang Cheng, Jiaying Liu
TL;DR
Existing low-light enhancement methods may amplify noise or lose details through subsequent denoising. This paper proposes sequential Retinex decomposition with noise-aware smoothing and gradient weighting, and reports effective joint enhancement and denoising across varied images.
Problem
Low-light enhancement methods can amplify inherent noise, while extra denoising procedures may ruin image details.
Method
The method sequentially estimates illumination and reflectance using spatial smoothness and weight matrices to suppress noise and improve contrast.
Results
Extensive experiments report effective joint low-light enhancement and denoising, with strong advantages over compared methods in both tasks.
Takeaways & Limitations
The approach produces enhanced low-light images that preserve local details while limiting noise in the reported experiments.
Abstract
from arXiv · showhide
Many low-light enhancement methods ignore intensive noise in original images. As a result, they often simultaneously enhance the noise as well. Furthermore, extra denoising procedures adopted by most methods ruin the details. In this paper, we introduce a joint low-light enhancement and denoising strategy, aimed at obtaining well-enhanced low-light images while getting rid of the inherent noise issue simultaneously. The proposed method performs Retinex model based decomposition in a successive sequence, which sequentially estimates a piece-wise smoothed illumination and a noise-suppressed reflectance. After getting the illumination and reflectance map, we adjust the illumination layer and generate our enhancement result. In this noise-suppressed sequential decomposition process we enforce the spatial smoothness on each component and skillfully make use of weight matrices to suppress the noise and improve the contrast. Results of extensive experiments demonstrate the effectiveness and practicability of our method. It performs well for a wide variety of images, and achieves better or comparable quality compared with the state-of-the-art methods.
I. INTRODUCTION
Low-light images commonly suffer from poor definition, low contrast, and noise, while existing enhancement strategies can amplify noise, over- or under-enhance images, or lose details. The proposed method jointly enhances and denoises through sequential Retinex decomposition.
- Low-light photographs often have low definition, low contrast, and substantial noise, motivating post-processing enhancement.
- Direct illumination amplification can saturate bright areas and lose details, while histogram equalization may under- or over-enhance images.
- Dehazing-based approaches and enhancement pipelines using BM3D or joint-bilateral filtering address low-light noise after enhancement.
- Retinex methods decompose an image into reflectance and illumination, but weighted variational estimation can leave noise especially visible in noisy inputs.
- The proposed method estimates illumination first, refines reflectance using the illumination and original image, and generates the result with Gamma-corrected illumination.
A. Overview
The overview extends the Retinex model with an explicit noise term and identifies how common estimation procedures retain or propagate noise. It then motivates a sequential optimization that separately estimates reflectance and illumination while accounting for noise.
- The modified Retinex formulation models the observed image as reflectance multiplied by illumination plus noise.
- Using R′ = S/L leaves most unpleasant noise in the reflectance because R′ = R + N/L.
- Extra denoising can produce noisy results while risking the loss of critical image details.
- Iterative simultaneous estimation allows noise in reflectance to impair the expected illumination map during updates.
- The proposed optimization calculates reflectance and illumination more separately while treating noise as an affecting factor.
B. The Sequential Estimation
The sequential estimation framework refines illumination and reflectance with fidelity, smoothness, and gradient constraints. Image-dependent weight matrices regulate reflectance smoothing and gradient adjustment to limit noise and strengthen contrast.
- Sequential equations estimate an undisturbed illumination and preferred reflectance in sequence.
- The illumination objective combines fidelity to the initial map with total-variation sparsity to encourage piece-wise smoothness.
- Reflectance estimation enforces fidelity to the observed image and accommodative spatial smoothness controlled by the weight matrix W.
- The gradient term minimizes the difference between reflectance and observed-image gradients, strengthening the contrast of the final result.
- For color images, all three channels share one illumination map, initialized from the input image's Y channel.
- The weight-controlled reflectance smoothness is intended to remain effective for intensive, large-scale noise in the original image.
- The adjusted gradient matrix G amplifies reflectance while restraining noise; λ controls amplification, σ its gradient-dependent rate, and ε filters small gradients.
- For each observed image, the matrices ˆL, W, and G need to be calculated only once.
C. The Solution
The method estimates illumination and reflectance sequentially, using smoothed illumination and adjusted gradients to suppress noise before gamma-corrected image reconstruction.
- The method estimates illumination independently before refining reflectance using both the observed image and illumination estimate.
- A weighted approximation replaces the illumination smoothness term with quadratic terms that can be solved directly.The formulation uses adjusted gradient weights derived from the estimated illumination.
- The illumination solution is obtained using matrix operators built from discrete gradient and diagonal weighting constructions.D contains forward-difference Toeplitz matrices, while vectorization and Diag(x) support the linear system.
- Figure 2 compares input, illumination, reflectance, result, and detail images for LIME and the proposed method.Rows are ordered with LIME on top and the proposed method below.
- After illumination and reflectance estimation, gamma correction adjusts illumination and combines it with reflectance to produce the enhancement result.The gamma parameter is empirically set to 2.2.
III. EXPERIMENTAL RESULTS
Experiments compare the method with established enhancement approaches on ordinary and noisy low-light images, emphasizing detail preservation and joint denoising.
- Experiments use MATLAB R2017a and test images from datasets provided by the authors of references and.Parameters are empirically fixed for the reported experiments.
- In noisy-image comparisons, all competing methods receive BM3D afterward, whereas the proposed method performs enhancement and denoising jointly.
- Figure 2 shows that the proposed method preserves local details, removes noise, and produces more detailed illumination and colorful reflectance maps than LIME.
- Compared with HE, SRIE, NPEA, and LIME, the proposed method better enhances the examples while retaining grass details that other methods lose or obscure.The reported comparisons note noise in HE, SRIE, and NPEA, and bright-area detail loss with LIME.
- The proposed method is reported to have strong advantages in both low-light enhancement and denoising compared with the evaluated alternatives.The alternatives are described as having difficult or less effective post-denoising, or detail loss and blurring.
IV. CONCLUSION
The paper addresses noise in low-light enhancement by presenting a sequential-decomposition method that jointly enhances images and suppresses noise.
- The paper identifies noise as an existing problem in mainstream low-light enhancement methods and proposes a joint method based on sequential decomposition.
- Extensive experimental results demonstrate the effectiveness of the proposed method.