Source-linked AI summary
Progressive Retinex: Mutually Reinforced Illumination-Noise Perception Network for Low Light Image Enhancement
Yang Wang, Yang Cao, Zheng-Jun Zha, Jing Zhang, Zhiwei Xiong, Wei Zhang, Feng Wu
TL;DR
Low-light enhancement must handle coupled contrast enhancement and noise suppression, since illumination and noise influence one another. This paper uses progressive mutual perception with pointwise networks and camera-model-based data synthesis; it reports superior performance on synthetic and real images.
Problem
Retinex-based enhancement involves contrast enhancement and noise suppression, but existing methods often treat them separately despite their coupling.
Source: §1 · p. 2Method
A progressive Retinex framework mutually models illumination and noise with two fully pointwise networks, alongside camera-imaging-model-based data synthesis.
Source: §1 · p. 2Results
Takeaways & Limitations
The pointwise framework is reported to suppress tiny-texture/noise ambiguity and improve computational efficiency.
Source: §1 · p. 2, §5 · p. 8The method captures statistical distributions in pixel space but neglects structural properties.
Source: §5 · p. 8
Abstract
from arXiv · showhide
Contrast enhancement and noise removal are coupled problems for low-light image enhancement. The existing Retinex based methods do not take the coupling relation into consideration, resulting in under or over-smoothing of the enhanced images. To address this issue, this paper presents a novel progressive Retinex framework, in which illumination and noise of low-light image are perceived in a mutually reinforced manner, leading to noise reduction low-light enhancement results. Specifically, two fully pointwise convolutional neural networks are devised to model the statistical regularities of ambient light and image noise respectively, and to leverage them as constraints to facilitate the mutual learning process. The proposed method not only suppresses the interference caused by the ambiguity between tiny textures and image noises, but also greatly improves the computational efficiency. Moreover, to solve the problem of insufficient training data, we propose an image synthesis strategy based on camera imaging model, which generates color images corrupted by illumination-dependent noises. Experimental results on both synthetic and real low-light images demonstrate the superiority of our proposed approaches against the State-Of-The-Art (SOTA) low-light enhancement methods.
1 INTRODUCTION
Low-light enhancement must address contrast enhancement and noise suppression together because illumination and noise are coupled. The paper proposes a progressive Retinex framework that models both through pointwise networks and reports superiority on synthetic and real low-light images.
1 INTRODUCTION
- Existing Retinex methods often treat contrast enhancement and noise suppression separately, although illumination affects noise and noise affects illumination statistics.Ignoring this coupling can produce inaccurate estimates and under- or over-smoothed enhancement results.Source: §1 · p. 2
- The proposed progressive Retinex framework mutually reinforces illumination and noise perception using two fully pointwise networks as statistical constraints.The two modeling processes proceed until stable estimates are obtained, with each preceding process benefiting from gradual improvement in the other.Source: §1 · p. 2
- The framework is reported to suppress ambiguity between tiny textures and noise while improving computational efficiency.The paper also proposes a camera-imaging-model-based data generation strategy to address insufficient training data.Source: §1 · p. 2
- The method is reported to outperform existing Retinex and other state-of-the-art low-light enhancement methods on synthetic and real data.The supplied introduction reports this comparison without quantitative values.Source: §1 · p. 2
2 RELATED WORK
Prior work includes model-based methods using Retinex decompositions and learning-based methods that learn illumination and reflectance regularities. The paper identifies the missing coupling between illumination and noise as a source of inaccurate estimates and proposes a progressive, pointwise alternative.
2 RELATED WORK
- Model-based methods decompose images into reflectance and illumination, an ill-posed inverse problem commonly addressed with priors or assumptions.Examples include smoothness and multi-scale assumptions for illumination estimation.Source: §2 · p. 2
- Learning-based methods learn illumination and reflectance regularities from training data to improve Retinex generalization.Examples include networks that learn illumination and noise separately or learn contrast-enhancement and denoising regularities simultaneously.Source: §2 · p. 3
- Existing methods may denoise after enhancement, but neglecting illumination-noise coupling can cause inaccurate estimates and under- or over-smoothing.The cited approaches include superpixel-adaptive BM3D, BM3D after illumination removal, and progressive bilateral filtering.Source: §2 · p. 3
- The proposed progressive framework uses fully pointwise convolutional units to select representative pixels for illumination and noise perception.The paper describes this representation as suppressing interference from tiny-texture/noise ambiguity and improving computational efficiency.Source: §2 · p. 3
3 PROPOSED METHOD
The method models illumination and noise statistics from representative pixels, then estimates them progressively so each estimate guides the other. Its network components use pointwise convolutions, and mean squared error supervises training.
3.1 Motivation
- Retinex enhancement is treated as illumination and noise-level estimation, with regularized constraints representing local statistics of ambient illumination and image noise.Representative illumination pixels are described as high-reflectance regions, while low-light noise pixels differ in color-channel intensity distribution from their neighborhoods.Source: §3.1 · pp. 3–4
3.2 Network Architecture
- Pointwise convolutions and max pooling select representative pixels by their RGB intensity distributions, whose statistics serve as Retinex regularities.IM-Net estimates illumination from representative pixels, while NM-Net estimates noise variance; patch variance can overestimate noise because it includes texture variance.Source: §3.1 · p. 4, §3.2 · pp. 4–5
- The progressive framework comprises IM-Net for illumination estimation and NM-Net for noise-level estimation, with a fully pointwise unit modeling pixel statistics.The method also uses guided image filtering to smooth the illumination map and BM3D with the estimated noise-level map during testing.Source: §3.2 · pp. 4–5
- Illumination and noise estimates are refined in stages, with each map guiding the other network to suppress responses from pixels relevant to the other task.The first IM-Net iteration uses a noise-level input set to zero; the process continues until results stabilize.Source: §3.2 · p. 5
3.3 Loss Function
- Mean squared error supervises the network, comparing its output with the ground truth across training samples.The passage defines n as the number of training samples, X_i as network output, and Y_i as ground truth.Source: §3.3 · p. 5
4 EXPERIMENTS
The experiments test synthetic-data generation, progressive-model choices, and comparisons with other low-light enhancement methods. The authors report strong benchmark performance and effective contrast enhancement and denoising, alongside high computational efficiency.
4.2 Synthetic Data Generation
- The synthesis strategy models illumination-dependent and independent noise through a camera-imaging process, producing 250,000 low-light patches for training and testing.The patches come from 200 well-exposed images; 200,000 are used for training and 50,000 for testing.Source: §4.2 · p. 5, Figure 5 · p. 6
4.1 Experimental Setup
- The IM-Net and NM-Net are trained for 50,000 iterations; ablations use 50 non-overlapping synthetic images and the same 15-layer residual base model.Training uses stochastic gradient descent with batch size 128 and initial learning rate 0.005.Source: §4.1 · p. 5, §4.3 · p. 6
4.3 Ablation Studies
- PSNR and SSIM improve as iterations increase and stabilize at the fourth iteration, which is also the reported stable point for the real-scene example.The four-iteration model is used as the default in the following experiments.Source: §4.3 · p. 6, Table 3 · p. 6, Figure 6 · p. 6
- Using intermediate enhancement or denoising results as network inputs performs worse as iterations increase because of accumulation error.The proposed setup instead passes the previous iteration’s noise or illumination estimate with the original patch and performs enhancement and denoising only at the final iteration.Source: §4.3 · p. 6, Table 4 · p. 6
- Denoising improves results over IM-Net alone, but a constant noise level performs worse than guidance from NM-Net..Source: §4.3 · pp. 6–7, Table 5 · p. 6
4.4 Comparisons with SOTA Methods
- The method achieves lower NIQE and higher PSNR/SSIM than other methods on IP100 and FNF38, and the best results in all three reported terms on LOL and NPE.On MPI, it matches the comparison method at PSNR 19.77 and has FSIM 0.9456 versus 0.9347.Source: Table 6 · p. 7, Table 7 · p. 7, §4.4 · p. 7
- On noisy real-scene images, the method removes noise while enhancing contrast without the over- or under-denoising reported for other approaches.The authors also report that competing methods can leave residual noise, while some introduce color cast, blur, or over-exposure.Source: Figure 8 · p. 8, §4.4 · p. 8
- The framework processes a 129*129*3 image in 0.46s on CPU and 0.03s with GPU acceleration, compared with 26.47s for the CNN-based method.The authors attribute this efficiency to the fully pointwise convolutional structure.Source: §4.4 · p. 8
5 CONCLUSION
The paper concludes that its progressive Retinex framework performs well on synthetic and real low-light images. It also identifies pixel-space-only statistics, which omit structural properties, as a limitation.
5 CONCLUSION
- The progressive framework uses fully pointwise convolutions to reduce tiny-texture/noise ambiguity and achieve high computational efficiency, with evaluations showing superior performance over representative SOTA methods.The evaluations cover synthetic and real low-light images.Source: §5 · p. 8
- The method captures pixel-space statistical distributions but neglects structural properties; a multi-branch network is proposed as future work to address both.The authors describe this as the method’s limitation and leave the proposed direction for future work.Source: §5 · p. 8