FLARE-BO: Fused Luminance and Adaptive Retinex Enhancement via Bayesian Optimisation for Low-Light Robotic Vision

arXiv cs.CV / 4/27/2026

📰 NewsTools & Practical UsageModels & Research

共有:

Key Points

The paper addresses the challenge of reliable low-light visual perception for autonomous robots by improving image enhancement without training a learned model.
It extends a prior Bayesian-optimisation-based, training-free approach by jointly tuning eight image-processing parameters across gamma correction, illumination normalization (LIME-style), chrominance denoising, multiple denoisers/filters, and Grey-World white balance.
The Bayesian optimisation setup uses unit hypercube parameter normalization, objective standardization, Sobol quasi-random initialization, and a Log Expected Improvement acquisition strategy to explore the larger parameter space effectively.
Experiments on the LOL (Low Light paired) dataset show that FLARE-BO achieves clear performance gains over existing methods, including those not specifically trained on the LOL dataset.
The framework is intended to reduce typical low-light artifacts by improving luminance/illumination handling and denoising behavior, mitigating issues like edge oversmoothing seen with NLM under noise.

Abstract

Reliable visual perception under low illumination remains a core challenge for autonomous robotic systems, where degraded image quality directly compromises navigation, inspection, and various operations. A recent training free approach showed that Bayesian optimisation with Gaussian Processes can adaptively select brightness, contrast, and denoising parameters on a per-image basis, achieving competitive enhancement without any learned model. However, that framework is limited to three parameters, applies no illumination decomposition or white balance correction, and relies on Non-Local Means denoising, which tends to over smooth edges under noisy conditions. This paper proposes FLARE-BO (Fused Luminance and Adaptive Retinex Enhancement via Bayesian Optimisation), an extended framework that jointly optimises eight parameters spanning across gamma correction, LIME-style illumination normalisation, chrominance denoising, bilateral filtering, NLM denoising, Grey-World automatic white balance, and adaptive post smoothing. The search engine employs a unit hypercube parameter normalisation, objective standardisation, Sobol quasi-random initialisation, and Log Expected Improvement acquisition for principled exploration of the expanded space. Performance of the proposed method is benchmarked using the Low Light paired dataset (LOL) and results show marked improvements of the proposed method over existing methods that were not specifically trained using this dataset.