Poisson Informed Retinex Network for Extreme Low-Light Image Enhancement

TL;DR

This work tackles extreme low-light image enhancement under Poisson noise by introducing PIRL, a lightweight encoder-decoder that jointly learns Retinex-based illumination and reflectance along with a Poisson noise component. The model enforces end-to-end decomposition through a Poisson-aware loss that combines reconstruction, Retinex estimation, and noise terms, with outputs for $L$, $R$, and $N$ and activation schemes to stabilize each component. By modeling the image as $Y = L \circ R \circ N$ with $X = L \circ R$, PIRL achieves improved color constancy and reduced noise without requiring priors on illumination or reflectance, outperforming several baselines in qualitative and competitive quantitative metrics on the LOL dataset. The method holds promise for real-time, low-power deployment and can be extended to raw sensor processing to further improve efficiency and fidelity under Poisson-dominated low-light conditions.

Abstract

Low-light image denoising and enhancement are challenging, especially when traditional noise assumptions, such as Gaussian noise, do not hold in majority. In many real-world scenarios, such as low-light imaging, noise is signal-dependent and is better represented as Poisson noise. In this work, we address the problem of denoising images degraded by Poisson noise under extreme low-light conditions. We introduce a light-weight deep learning-based method that integrates Retinex based decomposition with Poisson denoising into a unified encoder-decoder network. The model simultaneously enhances illumination and suppresses noise by incorporating a Poisson denoising loss to address signal-dependent noise. Without prior requirement for reflectance and illumination, the network learns an effective decomposition process while ensuring consistent reflectance and smooth illumination without causing any form of color distortion. The experimental results demonstrate the effectiveness and practicality of the proposed low-light illumination enhancement method. Our method significantly improves visibility and brightness in low-light conditions, while preserving image structure and color constancy under ambient illumination.

Figures (5)

• Figure 1: Proposed architecture for Poisson noise removal and low-light image enhancement. The network is structured with a convolutional encoder-decoder framework. The three branches of the output layers generate an illumination map, a reflectance map, and a noise map.
• Figure 2: Schematic of the loss function in our proposed method Poisson Informed Retinex of Low-light images (PIRL).
• Figure 3: Visual comparison of low-light enhancement results for Scene 1: AC image. We compare with recent low-light enhancement methods: (a) low-light input image, (b) LIME, (c) deep-retinex, and (e) Proposed. Notice that our method achieves better quality in the enhanced image than the other methods shown here.
• Figure 4: Plot of histograms of RGB channels of recovered AC image at low-light level 3 in Figure \ref{['fig:visual_ac']}. An image with good color consistency exhibits very similar histograms for RGB colors. Notice that our method gives the best color consistency resulting no color distortion unlike other methods.
• Figure 5: Visual comparison of low-light enhancement results on the Door image:: (a) low-light input image, (b) LIME, (c) deep-retinex, (d) zero-DCE, and (e) Proposed method PIRL. Our method achieves significantly better visual quality in the enhanced image than the other methods shown.