DLEN: Dual Branch of Transformer for Low-Light Image Enhancement in Dual Domains

TL;DR

The paper tackles low-light image enhancement (LLE) by introducing the Dual Light Enhance Network (DLEN), which fuses spatial and frequency-domain information via a learnable wavelet transform and a dual-branch Transformer-based restorer. DLEN comprises a Light Component Predictor (LCP), a Learnable Wavelet Network (LWN), and two attention-driven branches—Illumination Learning Branch (MIAB-based) and Structure Enhancement Branch (SEAB-based)—culminating in a fused output that preserves edges, textures, and natural color. The authors provide extensive experiments on LOLv1 and LOLv2 demonstrating state-of-the-art PSNR/SSIM gains and qualitative improvements, with code available at the project repository. Overall, DLEN advances LLE by explicitly modeling both illumination and structure in a frequency-aware, Transformer-based framework, offering practical gains for downstream vision tasks in challenging lighting conditions.

Abstract

Low-light image enhancement (LLE) aims to improve the visual quality of images captured in poorly lit conditions, which often suffer from low brightness, low contrast, noise, and color distortions. These issues hinder the performance of computer vision tasks such as object detection, facial recognition, and autonomous driving.Traditional enhancement techniques, such as multi-scale fusion and histogram equalization, fail to preserve fine details and often struggle with maintaining the natural appearance of enhanced images under complex lighting conditions. Although the Retinex theory provides a foundation for image decomposition, it often amplifies noise, leading to suboptimal image quality. In this paper, we propose the Dual Light Enhance Network (DLEN), a novel architecture that incorporates two distinct attention mechanisms, considering both spatial and frequency domains. Our model introduces a learnable wavelet transform module in the illumination estimation phase, preserving high- and low-frequency components to enhance edge and texture details. Additionally, we design a dual-branch structure that leverages the power of the Transformer architecture to enhance both the illumination and structural components of the image.Through extensive experiments, our model outperforms state-of-the-art methods on standard benchmarks.Code is available here: https://github.com/LaLaLoXX/DLEN

Figures (6)

• Figure 1: The target image (left) and the enhanced result (right) produced by our method. Our approach effectively enhances the image in color space.
• Figure 2: Visualized results on a benchmark dataset. Each image shows the input image, the output produced by our method, and the corresponding target image.
• Figure 3: The figure illustrates the detailed structure of our model, which consists of two main components: (a) the Light Component Predictor and (b) the Dual-Branch Restorer.
• Figure 4: The figure shows the qualitative experimental results on LOLv1. Our method effectively reduces color distortions and enhances lighting effects.
• Figure 5: This figure demonstrates the detailed effects of the module on the image. When the LWN module (a) is removed, there is a noticeable loss of texture preservation. Similarly, excluding the SEAB branch (b) results in a loss of crucial structural information. Our method (c), however, produces results that are visually closest to the target (d), with the most accurate preservation of both texture and structural details.