💻 computer science

RL-AWB: Deep Reinforcement Learning for Auto White Balance Correction in Low-Light Night-time Scenes

The paper presents RL-AWB, a novel framework that combines a statistical nighttime illuminant estimation algorithm with deep reinforcement learning to dynamically optimize auto white balance parameters without ground-truth data, alongside the introduction of a new multi-sensor nighttime dataset to validate its strong generalization across diverse lighting conditions.

Original authors: Yuan-Kang Lee, Kuan-Lin Chen, Chia-Che Chang, Yu-Lun Liu

Published 2026-07-09

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Yuan-Kang Lee, Kuan-Lin Chen, Chia-Che Chang, Yu-Lun Liu

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you're taking a photo of a friend at a dimly lit party. The camera's automatic "white balance" tries to guess what the light source is so your friend's skin looks natural, not blue or orange. In broad daylight, this is easy. But at night? It's a disaster. The light is weak, the camera gets noisy (like static on an old TV), and the automatic settings often get it completely wrong, leaving your photo looking like a weird, color-shifted mess.

The authors of this paper, RL-AWB, decided to fix this by teaching a computer to act like a seasoned photo editor who knows exactly how to tweak the knobs for every single night scene.

The Problem: The "One-Size-Fits-None" Trap

Most existing methods are like a chef who uses the exact same recipe for every dish, regardless of the ingredients. They rely on fixed rules or massive databases of training data. The paper argues that these approaches fail in the dark because the "rules" of how light behaves break down when it's too dark and noisy. Furthermore, if you train a model on one type of camera (say, an iPhone), it often gets confused when you switch to another (like a Sony), producing wild color errors. The paper explicitly rules out the idea that simply throwing more data at a deep learning model solves this; in fact, they found that standard deep learning models struggle to generalize across different cameras in low light.

The Solution: A Smart, Step-by-Step Detective

Instead of guessing the answer in one giant leap, the authors created a system called RL-AWB (Reinforcement Learning for Auto White Balance). Think of this system as a detective solving a mystery, but instead of clues, it's looking at the colors in the photo.

The Base Tool (SGP-LRD): First, they built a new statistical tool called SGP-LRD. Imagine this as a magnifying glass that looks for "gray pixels" (pixels that should be neutral gray) in the photo. In the dark, noise makes it hard to find these. This tool is smart: it filters out the noisy "fake" gray pixels and focuses on the reliable ones, using a special math trick (Minkowski norm) to weigh the evidence.
The Detective (The RL Agent): Here is where the magic happens. They didn't just let the tool run once. They trained a "Reinforcement Learning" agent—a digital learner that plays a game.
- The Game: The agent looks at the photo and the current settings of the tool.
- The Move: It decides to tweak two specific knobs: how many gray pixels to look for, and how strictly to weigh them.
- The Reward: If the tweak makes the colors look more accurate (measured by how much the "angular error" drops), the agent gets a point. If it makes things worse, it loses points.
- The Strategy: The agent doesn't just guess once. It takes N steps (usually around 3 iterations). It makes a small adjustment, checks the result, makes another small adjustment, and so on, until it finds the perfect setting for that specific image.

The "Training Camp" (Curriculum Learning)

Teaching this agent is tricky. If you throw it into a chaotic night scene immediately, it gets confused. So, the authors used a "curriculum learning" approach.

Stage 1: They taught the agent on just one image until it learned how to stop when it found a good answer.
Stage 2: Then, they introduced a small "class" of 5 images at a time. The agent practiced on these, cycling through them, learning to adapt quickly to different lighting conditions without needing thousands of examples.

The New Playground: The LEVI Dataset

To prove this works, the authors realized the existing night-time photo datasets were too small and only used one camera type. So, they created LEVI (Low-light Evening Vision Illumination).

This is the first multi-camera nighttime dataset.
It contains 700 images taken with two very different cameras: an iPhone 16 Pro and a Sony ILCE-6400.
The ISO (sensitivity) ranges from 500 to 16,000, covering everything from twilight to pitch-black darkness.
Every image has a "Color Checker" (a chart of known colors) to provide the ground truth for what the colors should be.

The Results: Fast, Accurate, and Adaptable

When they tested their method, the results were impressive, but the paper is careful to frame them as strong evidence rather than a magic bullet.

Accuracy: On the new LEVI dataset, their method achieved a median angular error of 3.08°, which is better than other statistical methods.
Data Efficiency: This is the big win. While other deep learning models needed to be trained on the entire dataset to work well, RL-AWB achieved similar or better results after training on just 5 images per dataset.
Cross-Sensor Magic: When they trained on the iPhone data and tested on the Sony data (and vice versa), RL-AWB held its ground. Other methods saw their error rates skyrocket (e.g., jumping from ~3° to over 13° or even 28° for some models). RL-AWB stayed low, around 3.03° and 1.99° respectively. This suggests the method is robust to different camera sensors.
Daytime Bonus: Even though it was built for night, they tested it on a daytime dataset (Gehler-Shi) and found it still performed very well, with a median error of 2.24°.

The Catch: When It Over-Corrects

The paper is honest about its limitations. Sometimes, if the initial guess is already pretty good, the agent gets a little too excited and "over-corrects," making the image slightly worse than it started. They call this a "failure mode" where the agent increases the error from an initial low estimate. They suggest future work will focus on making the agent more cautious in these specific scenarios.

The Bottom Line

The paper suggests that by combining a smart statistical tool with a learning agent that tweaks settings step-by-step, we can fix white balance in the dark much better than before. It works fast (about 1.1 seconds per image on a high-end GPU), needs very little training data, and doesn't get confused when you switch cameras. It's not a perfect, solved problem, but it's a significant step forward in teaching computers to see the night clearly.

Technical Summary: RL-AWB

Problem Statement

Automatic White Balance (AWB) is a critical component of camera image signal processing (ISP) pipelines, responsible for estimating scene illuminants and correcting color casts. While existing methods perform robustly in well-lit daytime scenarios, nighttime environments present fundamental challenges. Low illumination, high ISO settings, and severe chromatic noise violate the statistical assumptions underlying conventional AWB algorithms, leading to unstable illuminant estimates. Furthermore, deep learning-based approaches, though effective in daytime, often require extensive labeled training data and suffer from severe generalization degradation when deployed on unseen camera sensors or under different lighting conditions. The core difficulty lies in the breakdown of reliable color statistics in low-light scenes and the sensitivity of algorithmic parameters to specific scene characteristics, making cross-sensor adaptation particularly difficult.

Methodology

The authors propose RL-AWB, a novel framework that integrates Deep Reinforcement Learning (DRL) with a statistical color constancy algorithm to address nighttime AWB. The approach formulates AWB as a sequential decision-making problem where an RL agent dynamically selects optimal parameters for a statistical estimator based on the input image, without requiring ground-truth illuminants or reference images.

1. Core Algorithm: SGP-LRD

The foundation of the framework is SGP-LRD (Salient Gray Pixels with Local Reflectance Differences), a nighttime-specific statistical algorithm.

Salient Gray Pixel Detection: The algorithm models the linear image under a narrow-band spectral assumption. It identifies achromatic (gray) pixels by analyzing local contrast vectors in the logarithmic domain. Pixels are selected based on their alignment with the gray direction.
Refinement Filters: To mitigate noise and chromatic outliers prevalent in low-light images, two filtering stages are applied:
- Local Variance Filtering: Removes pixels with insufficient color information (low intra-pixel variance).
- Color Deviation Filtering: Removes pixels that deviate significantly from the scene's dominant color cast.
Local Reflectance Difference: The algorithm computes a pixel-wise normalized local reflectance difference within overlapping 3x3 windows. This design amplifies high-SNR gray signals through spatial redundancy while suppressing isolated spurious noise pixels.
Illuminant Estimation: The final illuminant is estimated using a Minkowski norm ( $p$ ) that accumulates weighted SGP intensities and normalized local differences. The algorithm relies on two tunable hyperparameters: the gray-pixel sampling percentage ( $N$ ) and the Minkowski order ( $p$ ).

2. Reinforcement Learning Framework

The RL-AWB framework uses the Soft Actor-Critic (SAC) algorithm to learn adaptive policies for tuning the $N$ and $p$ parameters of SGP-LRD.

State Design: The state vector combines illumination-related features (a log-chrominance RGB-uv histogram) and parameter-related features (a history of recent parameter adjustments and a timestep counter). Crucially, the state does not include ground-truth illuminants, ensuring deployability.
Action Design: The agent outputs continuous relative actions to update $N$ and $p$ . Actions are sampled, squashed via tanh, and rescaled to valid ranges, allowing for smooth, coordinated parameter adjustments.
Reward Design: The reward signal is based on the relative improvement in angular error ( $E_0 - E_t$ ) compared to the initial error. To stabilize training across images with varying initial difficulties, the reward is normalized. A difficulty-aware relaxation penalizes large, unstable moves. Episodes terminate after three steps of estimation stability or upon reaching a convergence bonus.
Curriculum Learning: Training employs a two-stage curriculum:
1. Stage 1: Single-image parameter tuning to stabilize behavior and teach the agent to detect convergence.
2. Stage 2: Multi-image adaptive tuning using a cyclic schedule over a small pool of images (e.g., 5 images) to balance exploration and stability.

Key Contributions

SGP-LRD Algorithm: A novel statistical color constancy algorithm tailored for nighttime scenes that achieves state-of-the-art illuminant estimation on public benchmarks by leveraging spatial redundancy and noise-aware filtering.
RL-AWB Framework: The first application of deep reinforcement learning to nighttime color constancy. It enables adaptive, per-image parameter optimization with exceptional data efficiency, requiring only a few training images (5 per dataset) to generalize across unseen sensors.
LEVI Dataset: The introduction of LEVI (Low-light Evening Vision Illumination), the first multi-camera nighttime dataset comprising 700 linear RAW images from two distinct systems (iPhone 16 Pro and Sony ILCE-6400) with manual annotations. This dataset facilitates rigorous cross-sensor evaluation, addressing a gap in existing benchmarks like NCC which are single-sensor.

Experimental Results

Experiments were conducted on the NCC, LEVI, and Gehler-Shi datasets.

Statistical Baselines: SGP-LRD outperformed existing statistical methods (e.g., GE-1st, MSGP, GI) on both NCC and LEVI datasets, achieving the lowest median and mean angular errors.
Cross-Dataset Generalization: In cross-dataset evaluations (e.g., training on NCC, testing on LEVI, and vice versa), RL-AWB demonstrated superior robustness compared to fully-supervised deep learning baselines (such as FC4, C5, GCC). While deep learning models suffered significant performance degradation due to domain and sensor shifts, RL-AWB maintained competitive accuracy.
Data Efficiency: Despite being trained on only 5 images per dataset, RL-AWB generalized better than models trained on full datasets. This highlights the efficacy of using RL to tune interpretable statistical parameters rather than directly regressing illuminants.
Daytime Generalization: When adapted to the well-lit Gehler-Shi dataset, RL-AWB achieved state-of-the-art generalization capabilities, reducing median angular error by 5.9% compared to the base SGP-LRD algorithm.
Ablation Studies: Comparisons with one-step regressors (MLP and CNN) confirmed that the iterative refinement mechanism of RL-AWB allows for error correction and recovery from suboptimal intermediate decisions, outperforming single-step prediction models.

Significance and Claims

The paper claims that RL-AWB represents a significant step forward by successfully bridging the gap between the interpretability and sensor-agnostic nature of statistical methods and the adaptive capability of learning-based approaches. By formulating AWB as a sequential decision-making problem, the method avoids the heavy data requirements and generalization failures typical of purely deep learning approaches in low-light, cross-sensor scenarios. The authors emphasize that their method provides a practical, deployment-compatible solution that operates without ground-truth information at inference time while effectively handling the complex, noisy conditions of nighttime imaging. The introduction of the LEVI dataset is also highlighted as a critical contribution for enabling standardized cross-sensor evaluation in the field of nighttime color constancy.