Ultra-High-Definition Low-Light Image Enhancement: A Benchmark and Transformer-Based Method (2212.11548v1)

Abstract: As the quality of optical sensors improves, there is a need for processing large-scale images. In particular, the ability of devices to capture ultra-high definition (UHD) images and video places new demands on the image processing pipeline. In this paper, we consider the task of low-light image enhancement (LLIE) and introduce a large-scale database consisting of images at 4K and 8K resolution. We conduct systematic benchmarking studies and provide a comparison of current LLIE algorithms. As a second contribution, we introduce LLFormer, a transformer-based low-light enhancement method. The core components of LLFormer are the axis-based multi-head self-attention and cross-layer attention fusion block, which significantly reduces the linear complexity. Extensive experiments on the new dataset and existing public datasets show that LLFormer outperforms state-of-the-art methods. We also show that employing existing LLIE methods trained on our benchmark as a pre-processing step significantly improves the performance of downstream tasks, e.g., face detection in low-light conditions. The source code and pre-trained models are available at https://github.com/TaoWangzj/LLFormer.

• The paper introduces the UHD-LOL dataset and the transformer-based LLFormer model, setting a new standard for ultra-high-definition low-light image enhancement.
• The proposed LLFormer utilizes axis-based multi-head self-attention and a dual gated feed-forward network to efficiently enhance image quality under low-light conditions.
• Extensive experiments demonstrate LLFormer’s superior performance with higher PSNR and SSIM metrics, highlighting its practical applications in high-resolution imaging tasks.

Ultra-High-Definition Low-Light Image Enhancement: A Benchmark and Transformer-Based Method

This paper addresses the task of low-light image enhancement (LLIE), focusing specifically on ultra-high-definition (UHD) images, including resolutions of 4K and 8K. As advances in optical sensor technology enable the capture of UHD images, there arises a need for enhanced image processing techniques capable of handling these large-scale data efficiently. The primary contributions of this paper include the introduction of a novel benchmark dataset for UHD low-light image enhancement and the development of an effective transformer-based model for this task.

Benchmark Dataset: UHD-LOL

The paper presents the UHD-LOL dataset, which is the first large-scale benchmark specifically designed for UHD-LLIE. This dataset comprises two subsets: UHD-LOL4K and UHD-LOL8K, containing images of 4K and 8K resolution, respectively. The subsets include training and testing pairs of low-light and normal-light images, synthesized using a defined protocol to ensure the inclusion of realistic image characteristics.

Proposed Method: LLFormer

The research introduces a new transformer-based image enhancement architecture named LLFormer. The main novelty of the LLFormer lies in its design, which comprises several key components:

• Axis-Based Multi-Head Self-Attention (A-MSA): This mechanism reduces computational complexity by performing self-attention separately along height and width axes, thereby improving computational efficiency while preserving the ability to capture long-range dependencies.
• Cross-Layer Attention Fusion Block: This feature dynamically learns the attention weights across different layers' features and fuses them adaptively, enhancing feature representation while maintaining the hierarchical structure of the network.
• Dual Gated Feed-Forward Network (DGFN): By incorporating a dual gated mechanism, DGFN focuses on extracting useful features, thereby further enhancing the network's ability to improve low-light conditions.

Experimental Results

The paper provides extensive benchmarking results across several datasets, highlighting LLFormer’s superior performance compared to existing LLIE methods. It achieves state-of-the-art outcomes on the proposed UHD-LOL dataset, as well as on other public datasets like LOL and MIT-Adobe FiveK. LLFormer is demonstrated to outperform both traditional and recent transformer-based methods in terms of PSNR, SSIM, LPIPS, and MAE metrics.

Implications and Future Directions

The development of LLFormer and the comprehensive UHD-LOL dataset suggest significant implications for both practical applications and theoretical advancements:

1. Enhanced LLIE Algorithms: Developing efficient methods for UHD-LLIE is crucial as the demand for image processing in high-resolution continues to grow, particularly in fields such as surveillance and autonomous driving where low-light conditions frequently occur.
2. Facilitation of Downstream Tasks: The paper also explores the benefit of employing LLFormer as a pre-processing step for improving the performance of downstream tasks like face detection, showcasing its potential utility beyond simple image enhancement.
3. Future Research: The introduction of a UHD-standard benchmark provides a foundation for further development of LLIE methods tailored for high-resolution contexts. Future work could involve optimizing computational efficiency and exploring adaptive methods that respond to varying levels of low-light conditions in complex environments.

The paper offers a substantial contribution to the field of image enhancement by setting a new standard for evaluating LLIE methodologies and providing a robust framework for developing sophisticated enhancement algorithms. As such, it not only establishes practical benchmarks but also lays groundwork for continuous innovations in the enhancement of UHD low-light images.