Low-Light Image Enhancement with Illumination-Aware Gamma Correction and Complete Image Modelling Network

Yinglong Wang, Zhen Liu, Jianzhuang Liu, Songcen Xu, Shuaicheng Liu

Abstract

Capturing high-quality images in low-light environments remains a significant challenge in computer vision due to noise, low contrast, and the loss of fine-grained details. This paper introduces the Illumination-Aware Gamma Correction (IAGC) framework, a novel approach that combines learnable, adaptive gamma correction with a specialized Transformer architecture. By utilizing Taylor Series approximations for mathematical efficiency and a "local-to-global" attention mechanism, the IAGC framework effectively recovers hidden details in extremely dark regions, outperforming existing state-of-the-art methods in both accuracy and computational performance.

Introduction

Low-light image enhancement is essential for various applications, from surveillance to mobile photography. However, existing solutions often struggle with a fundamental trade-off. Convolutional Neural Networks (CNNs) typically suffer from "locality bias," failing to utilize global context to reconstruct distant dark regions. Conversely, Transformer-based methods, while capable of modeling long-range dependencies, often rely on downsampling or patch-level attention to manage computational costs, which can discard the fine-grained pixel information necessary for high-fidelity restoration.

Furthermore, traditional Gamma Correction—a standard technique for adjusting image brightness—is usually applied using fixed, empirical values. The IAGC framework addresses these limitations by making gamma correction an adaptive, learnable component of the network architecture, applied at both global and local scales.

Methodology

The IAGC framework employs a three-stage "coarse-to-fine" pipeline to restore low-light images:

1. Global Gamma Correction Module (GGCM)

The process begins with the GGCM, which analyzes the entire image to predict a single global gamma factor ($\Gamma_g$). This factor lifts the overall brightness of the frame. To ensure the model remains efficient during training and inference, the authors utilize a Taylor Series approximation of the exponential gamma function ($I^\Gamma$). This polynomial approximation avoids the high computational cost of backpropagating through exponential operations on GPUs.

2. COmpletely MOdelling Vision Transformer (COMO-ViT)

The central component of the architecture is the COMO-ViT, which implements a hierarchical attention mechanism. Unlike models that downsample images, COMO-ViT maintains pixel-level dependencies. It first models local dependencies within specific windows and then utilizes a global Transformer module to connect these windows. This "complete modeling" allows dark, low-information pixels to "borrow" context from informative, brighter regions across the entire image.

3. Local Gamma Correction Module (LGCM)

In the final stage, the model predicts a pixel-wise gamma map ($\Gamma_l$). This local adjustment fine-tunes the illumination for specific areas, correcting subtle color deviations or shadows that a global adjustment might overlook.

Experimental Results

The IAGC framework was evaluated using the LOL (Low-Light) dataset suite, including LOL-v1, LOL-v2-real, and LOL-v2-synthetic.

Quantitative Performance

On the LOL-v1 dataset, IAGC achieved state-of-the-art results, reaching a PSNR of 24.53 dB and an SSIM of 0.842. This performance surpasses previous leading models such as MIRNet (24.14 dB) and SNR.

Qualitative Analysis

Visual comparisons indicate that IAGC recovers significantly more detail in regions that appear completely black in the input. Furthermore, the model maintains superior "hue recovery," avoiding the yellowish or greyish tints often introduced by competing methods.

Efficiency

By integrating the Taylor Series trick, the authors achieved a significant boost in running efficiency. This allows the complex Transformer blocks to remain performant without sacrificing the accuracy of the restoration.

Conclusion and Future Work

The IAGC framework represents a major advancement in low-light image enhancement by integrating learnable illumination priors directly into a Transformer architecture. While the model sets new benchmarks, the authors note that extremely complex lighting gaps can still result in slight local color deviations. Future research will focus on refining these edge cases and integrating more robust color-consistency priors to ensure perfect hue restoration across all environments.

Summary updated

Abstract

This paper presents a novel network structure with illumination-aware gamma correction and complete image modelling to solve the low-light image enhancement problem. Low-light environments usually lead to less informative large-scale dark areas, directly learning deep representations from low-light images is insensitive to recovering normal illumination. We propose to integrate the effectiveness of gamma correction with the strong modelling capacities of deep networks, which enables the correction factor gamma to be learned in a coarse to elaborate manner via adaptively perceiving the deviated illumination. Because exponential operation introduces high computational complexity, we propose to use Taylor Series to approximate gamma correction, accelerating the training and inference speed. Dark areas usually occupy large scales in low-light images, common local modelling structures, e.g., CNN, SwinIR, are thus insufficient to recover accurate illumination across whole low-light images. We propose a novel Transformer block to completely simulate the dependencies of all pixels across images via a local-to-global hierarchical attention mechanism, so that dark areas could be inferred by borrowing the information from far informative regions in a highly effective manner. Extensive experiments on several benchmark datasets demonstrate that our approach outperforms state-of-the-art methods.
BibTeX
@misc{wang2023low,
  author = {Yinglong Wang and Zhen Liu and Jianzhuang Liu and Songcen Xu and Shuaicheng Liu},
  title = {{Low-Light Image Enhancement with Illumination-Aware Gamma Correction and Complete Image Modelling Network}},
  year = {2023},
  eprint = {2308.08220},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2308.08220},
}
Figure