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

2026-07-08 · Source: cs.CV updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Computer Vision & Image Processing · Depth: Expert, extended

Summary

RL-AWB is a new framework designed to address the challenging problem of automatic white balance (AWB) correction in low-light nighttime scenes. It integrates a novel statistical algorithm, SGP-LRD, with a deep reinforcement learning (DRL) approach. SGP-LRD is tailored for nighttime conditions, employing salient gray pixel detection and a new illumination estimation method. The DRL component, RL-AWB, is the first of its kind for color constancy, using Soft Actor-Critic (SAC) and two-stage curriculum learning to dynamically optimize SGP-LRD's parameters (gray pixel selection threshold N% and Minkowski norm exponent p) for individual images. To facilitate evaluation, the authors introduce LEVI, the first multi-camera nighttime dataset, comprising 700 linear RAW images from iPhone 16 Pro and Sony ILCE-6400, with ISOs ranging from 500 to 16,000. Experiments demonstrate RL-AWB's superior generalization across low-light and well-illuminated images, achieving competitive performance with only 5 training images per dataset and strong cross-sensor robustness.

Key takeaway

For Computer Vision Engineers developing camera ISP pipelines, if you are struggling with robust auto white balance in low-light or cross-sensor environments, consider integrating deep reinforcement learning for dynamic parameter optimization. This RL-AWB framework, which adaptively tunes statistical algorithm parameters, significantly improves cross-sensor generalization and performance in challenging nighttime conditions. You can achieve superior accuracy with minimal training data, making it a practical solution for mobile photography and surveillance systems.

Key insights

Deep reinforcement learning can adaptively tune statistical auto white balance parameters for robust nighttime color constancy.

Principles

• Hybrid statistical-learning models offer both interpretability and adaptive optimization.
• AWB parameter tuning can be framed as a sequential decision-making problem for DRL.
• Curriculum learning improves DRL data efficiency and training stability for image processing tasks.

Method

The RL-AWB framework uses a Soft Actor-Critic agent to dynamically adjust two SGP-LRD parameters (gray-pixel sampling percentage N% and Minkowski order p) based on image log-chrominance histograms and parameter history, guided by a relative error improvement reward and two-stage curriculum learning.

In practice

• Implement dual-branch MLP encoders for RL states combining high-dimensional image features and low-dimensional history.
• Utilize relative, continuous actions for smooth and coordinated parameter updates in DRL-based image tuning.
• Create multi-camera datasets with color checker annotations to enable rigorous cross-sensor evaluation.

Topics

• Deep Reinforcement Learning
• Auto White Balance
• Color Constancy
• Nighttime Imaging
• Cross-Sensor Generalization
• LEVI Dataset

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, Computer Vision Engineer

Related on AIssential

• DUALVISION: RGB-Infrared Multimodal Large Language Models for Robust Visual Reasoning — Integrating infrared and RGB data via localized cross-attention significantly boosts MLLM robustness in degraded visual conditions.
• RL-ACRGNet: Reinforcement Learning-Based Chest Radiology Report Generation Network — RL-ACRGNet uses reinforcement learning and a dual-network approach to generate accurate, clinically coherent chest radiology reports.
• RL2ML: Finite-Rollout Surrogate Objectives from Reinforcement Learning to Maximum Likelihood — RL2ML offers unbiased, finite-rollout surrogate objectives connecting RL and ML, with optimal choice depending on metrics and variance.
• On Asymmetric Optimization of Reasoning and Perception in Vision-Language Model Post-Training — VLM post-training creates a perception-reasoning asymmetry due to token imbalance (SFT) or reward coupling (RL), hindering end-to-end performance.
• SSL-R1: Self-Supervised Visual Reinforcement Post-Training for Multimodal Large Language Models — SSL-R1 uses self-supervised visual puzzles to provide scalable, verifiable rewards for MLLM reinforcement learning.
• WorldSample: Closed-loop Real-robot RL with World Modelling — WorldSample uses a real-synthetic loop and Policy-Paced Learning to augment real-robot RL data, reducing interaction costs and improving performance.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CV updates on arXiv.org.