Source-linked AI summary
RENOIR - A Dataset for Real Low-Light Image Noise Reduction
Josue Anaya, Adrian Barbu
TL;DR
Artificially corrupted images may not accurately represent denoising performance on real low-light noise. The paper introduces a real-noise dataset with aligned images, estimates noise characteristics, and evaluates denoising methods, finding a Poisson-Gaussian trend and limitations in parameter estimation.
Problem
Artificial noise may not provide an accurate picture of denoising performance on real low-light noise.
Method
The paper constructs a dataset from low-light RAW images, develops pixel-aligned low- and high-noise image acquisition, and uses averaged clean images as ground-truth estimates.
Results
The noise follows an overall trend fitting the Poisson-Gaussian model, while the dataset evaluates denoising under real low-light noise.
Takeaways & Limitations
The dataset introduces an additional training and testing challenge compared with using image pairs.
Takeaways & Limitations
Wavelet-based estimation has difficulty estimating correct Poisson-Gaussian parameters because of short-range interactions in the noise.
Abstract
from arXiv · showhide
Image denoising algorithms are evaluated using images corrupted by artificial noise, which may lead to incorrect conclusions about their performances on real noise. In this paper we introduce a dataset of color images corrupted by natural noise due to low-light conditions, together with spatially and intensity-aligned low noise images of the same scenes. We also introduce a method for estimating the true noise level in our images, since even the low noise images contain small amounts of noise. We evaluate the accuracy of our noise estimation method on real and artificial noise, and investigate the Poisson-Gaussian noise model. Finally, we use our dataset to evaluate six denoising algorithms: Active Random Field, BM3D, Bilevel-MRF, Multi-Layer Perceptron, and two versions of NL-means. We show that while the Multi-Layer Perceptron, Bilevel-MRF, and NL-means with soft threshold outperform BM3D on gray images with synthetic noise, they lag behind on our dataset.
1. Introduction and Motivation
Real low-light noise is more complex than common artificial noise benchmarks, motivating a dataset with naturally noisy images and aligned low-noise counterparts for realistic evaluation.
- Motivation: Artificial-noise evaluations may not accurately represent denoising performance on real low-light camera images.Common benchmarks use small images with known, artificially added noise types, whereas low-light noise can have intensity-dependent variance and short-range correlations.
- Dataset: The paper introduces color images naturally corrupted by low-light noise from three cameras, including two Canon models and a Xiaomi mobile phone.The cameras are a Canon PowerShot S90, Canon EOS Rebel T3i, and Xiaomi Mi3.
- Evaluation: The paper evaluates noise estimation, the Poisson-Gaussian model, and six denoising algorithms using the collected images.The evaluation includes noise-level and PSNR estimation, model assessment, and denoising comparisons.
- Dataset: The dataset provides naturally noisy images and low-noise counterparts of the same scenes, with pixel and intensity alignment procedures.The authors describe short-exposure low-light images paired with long-exposure images of the same scene, alongside alignment methods.
- Significance: The dataset is intended to support realistic denoising evaluation, noise-structure analysis, and assessment of generalization across cameras.Its real low-light noise and varied camera sources address limitations of artificial Gaussian-noise databases.
2. Acquisition of Natural Image Pairs
The dataset contains natural low-light noise paired with low-noise images from the same static scenes, using a controlled acquisition procedure across three cameras. The authors address alignment and mobile-camera challenges while preserving substantial image and patch variability.
- Dataset composition: 120 low-light scenes provide noisy and low-noise color-image pairs, with approximately four images acquired per scene.Two images contain noise and two contain very little noise.
- Acquisition procedure: The acquisition keeps scenes static so that image differences primarily reflect noise rather than scene changes.The procedure places low-noise images before and after noisy captures and uses tripod-based or programmed acquisition where possible.
- Acquisition procedure: The sandwich procedure captures a low-noise reference, one or two higher-sensitivity noisy images, and a final low-noise clean image.The reference and clean images use low light sensitivity and long exposure, while noisy images use increased sensitivity and shorter exposure.
- Dataset composition: 40 scenes were collected for each of the Canon S90, Canon T3i, and Xiaomi Mi3 cameras.All images were collected in RAW format where the acquisition setup supported it.
- Dataset scale: Although the database contains fewer noisy images than some artificial-noise datasets, its images contain about 60 times more pixels and therefore more patch variability.This scale is presented as useful for studying noise models from one of the three cameras.
3. Dataset Information
The dataset analysis quantifies image fidelity and noise across camera batches using standard metrics and a noise-aware PSNR estimate. It also shows that nominally low-noise images retain measurable, visually imperceptible noise.
- The study quantifies image fidelity across batches using PSNR, SSIM, and VSNR.
- PSNR incorporates the proposed noise estimate rather than the standard clean–noisy difference-image estimate.
- PSNR and noise-level variation are summarized separately for each camera.
- The low-noise images retain noise levels of about σ = 3 and up to 5, although this noise is invisible to the eye.
- The authors suggest that perturbations with Gaussian noise of σ ≤5 can still produce natural-looking image patches.
- These observations may inform studies of natural-image statistics and generative models.
4. Experiments
The experiments evaluate alignment, noise estimation, the Poisson-Gaussian model, and denoising on synthetic and real low-light data. The proposed estimator is substantially more accurate than the standard difference-based estimate, while BM3D performs best on the real dataset.
- Noise estimation: The experiments compare the proposed noise estimator with the standard difference-image estimate on synthetic and real noise, and examine a Poisson-Gaussian model.
- Alignment and synthetic noise: The evaluation constructs aligned 8-bit reference, noisy, clean, and ground-truth images from 16-bit RAW data and artificial Gaussian noise.
- Alignment and synthetic noise: The alignment is identifiable only up to a multiplicative constant, so it is evaluated indirectly through noise-estimation quality.
- Noise estimation: For noisy images, the proposed estimator keeps relative error below 0.5%, versus around 5% for the standard estimate.
- Noise estimation: For low-noise images, the proposed method has error below 1%, compared with around 40% for the standard method.
- Noise estimation: Across 90 images, the relative estimation error is below 1% for all but four images.
- Poisson-Gaussian model: On the calibration scene, the proposed σ estimates are less biased and less variable, with average relative error 1.58% versus 36.22% for the standard method.
- Poisson-Gaussian model: The blurred reference and blurred noisy image closely approximate IGT, while box-kernel smoothing remains imperfect because it removes noise less effectively than Gaussian smoothing.
5. Conclusions
The paper introduces a real low-light noise dataset with aligned image pairs and evaluates noise estimation, noise modeling, and denoising algorithms on it. The results support the dataset’s value for quantitative analysis and expose limitations of current noise estimation and denoising methods under real noise.
- Dataset and alignment: The dataset contains real low-light noise and supports training or evaluating denoising algorithms on digital-camera images.Images were acquired under low-light settings from two digital cameras and a mobile phone.
- Dataset and alignment: Pixel-aligned RAW images and intensity-aligned BMP images make the noisy-clean difference attributable to noise rather than misalignment.The paper also presents a technique to calculate PSNR without a ground truth.
- Noise modeling and estimation: The noise follows an overall Poisson-Gaussian trend, but wavelet-based estimation struggles to recover model parameters.The difficulty is attributed to short-range noise interactions and using a box filter instead of a Gaussian filter for smoothing.
- Denoising evaluation: The dataset evaluates six denoising algorithms, including ARF, BM3D, opt-MRF, MLP, and two versions of Non-local Means.Noise levels in denoised images were assessed using PSNR, VSNR, and SSIM.
- Denoising evaluation: Methods trained or tuned on artificial Gaussian noise could perform better on low-light images if trained on this dataset.The paper therefore positions the dataset as a basis for proper analysis of denoising tools and quantitative evaluation of noise models.
- Denoising evaluation: Real low-light data pose a broader challenge than artificial-noise benchmarks because they contain a large range of noise levels.This variation complicates training and testing but helps denoising algorithms advance toward practical automatic camera-image denoising.