Event-Illumination Collaborative Low-light Image Enhancement with a High-resolution Real-world Dataset

Senyan Xu, Zhijing Sun, Kean Liu, Xin Lu, Ruixuan Jiang, Mingyang HuangXueyang Fu, Zheng-Jun Zha University of Science and Technology of China {syxu,sunzhijing}@mail.ustc.edu.cn, {xyfu,zhazj}@ustc.edu.cn · 2026

Read Event-Illumination Collaborative Low-light Image Enhancement with a High-resolution Real-world Dataset by Senyan Xu, Zhijing Sun, Kean Liu, Xin Lu, Ruixuan Jiang, Mingyang HuangXueyang Fu, Zheng-Jun Zha University of Science and Technology of China {syxu,sunzhijing}@mail.ustc.edu.cn, {xyfu,zhazj}@ustc.edu.cn (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2605-22186.txt.

Introduction

Senyan Xu and co-authors present EIC-LIE, an event-illumination collaborative framework for low-light image enhancement. Event-based methods exploit high dynamic range from events but overlook global illumination and event noise in real scenes. Their EICI module gathers HDR features across lighting conditions and injects complementary content into illumination and event representations. An illumination-aware event filter then cuts event noise using brightness statistics from images. The team also builds a beam-splitter-based hybrid imaging system for real-world data collection.

Catalog record

CREATORSenyan Xu, Zhijing Sun, Kean Liu, Xin Lu, Ruixuan Jiang, Mingyang HuangXueyang Fu, Zheng-Jun Zha University of Science and Technology of China {syxu,sunzhijing}@mail.ustc.edu.cn, {xyfu,zhazj}@ustc.edu.cn
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE34895
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2605.22186

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Event-Illumination Collaborative Low-light Image Enhancement with a High-resolution Real-world Dataset Senyan Xu, Zhijing Sun, Kean Liu, Xin Lu, Ruixuan Jiang, Mingyang HuangXueyang Fu, Zheng-Jun Zha University of Science and Technology of China {syxu,sunzhijing}@mail.ustc.edu.cn, {xyfu,zhazj}@ustc.edu.cn Abstract Event-based low-light image enhancement (LIE) methods mainly focus on incorporating high dynamic range (HDR) information from events while overlooking the essential global illumination in images and the inherent noise sensitivity of event signals in real-world scenarios. To address these issues, we propose EIC-LIE, an event-illumination collaborative LIE framework. Concretely, we first design an Event-Illumination Collaborative Interaction (EICI) module, which contains two key processes: forward gathering, which gathers HDR features across varying lighting conditions, and backward injection, which provides complementary content for illumination and event representations. Next, we introduce an Illumination-aware Event Filter (IAEF) that dynamically reduces event noise based on brightness statistics derived from images. Additionally, we build a beam-splitter-based hybrid imaging system to collect high-quality event-image pairs with temporal synchronization from dynamic scenes, providing the first high-resolution, real-world event-based LIE dataset. Extensive experiments show that our EIC-LIE outperforms state-of-the-art methods on five real-world and synthetic datasets, significantly surpassing previous methods with improvements of up to 1.24dB in PSNR and 0.069 in SSIM. The code and dataset are released at https://github.com/QUEAHREN/EIC-LIE. 1 Introduction In low-light environments, captured images often suffer from poor visibility, increased noise, and texture loss due to the limitations of traditional sensors. These issues not only impair human visual perception but also negatively impact high-level vision tasks such as semantic segmentation 80 , object detection 52 ; 62 , and tracking 68 . The rapid development of deep learning has significantly enhanced traditional image enhancement techniques 64 ; 77 ; 57 ; 16 . Specifically, recent low-light image enhancement (LIE) methods 4 ; 66 have achieved promising results by leveraging Retinex theory 27 . These methods incorporate illumination priors to guide the enhancement network, thereby improving the overall quality of the images. Figure 1 : Visual comparison of LIE methods on the proposed real-world RLE dataset. To better illustrate the enhancement effects, color invariants 15 are adopted as visualization tools. Specifically, the invariant C can be interpreted as describing object color regardless of intensity, while W functions as an edge detector specific to changes in spectral distribution. See supp. for more details regarding the invariants. Compared to traditional frame-based cameras, event cameras 78 ; 33 offer unique advantages in low-light conditions by capturing high dynamic range (HDR), high temporal resolution, and low-latency event streams. Nevertheless, current event-based LIE approaches still encounter multiple challenges: (1) Existing event-based LIE methods 36 ; 24 ; 30 heavily rely on direct feature fusion strategies, using complex fusion modules to compensate for missing image details. However, these methods often overlook the essential global illumination information emphasized in Retinex-based LIE techniques. (2) Event-based LIE is constrained by the inherent noise sensitivity of event signals in low-light conditions. As the number of available photons decreases, especially with low event-triggering thresholds ( c in Eq. 3 )— random noise significantly increases 9 . 30 propose selectively fusing events from high Signal-to-Noise Ratio (SNR) regions, using an SNR map estimated from images. However, this fixed-guidance strategy lacks reliability, leading to noticeable noise in the enhanced results, as demonstrated in Figure 1 . (3) The lack of high-resolution, high-quality, real-world datasets has significantly hindered the progress of event-based LIE (see Table 1 ). While SDE dataset 30 leverages a robotic arm to capture ground truth (GT) images along the same paths, their method is restricted to static scenes with slow camera movements, unavoidable temporal misalignment ( <10ms ), and severe color inaccuracies and low resolution due to the DAVIS346 sensor 3 . To address these challenges, we propose EIC-LIE, an Event-Illumination Collaborative Low-light Image Enhancement (LIE) framework. It consists of two key components, including Event-Illumination Collaborative Interaction (EICI) and Illumination-Aware Event Filter (IAEF) . The former enables collaborative bidirectional interaction between event and illumination information, merging high dynamic range (HDR) details with global illumination priors. The later dynamically filters event noise by utilizing bright statistics derived from RGB images. The visual analysis by introducing several physical invariants 15 has been shown in Figure 1 , EvLight 30 exhibits severe noise interference in invariant C , while MambaLLIE 66 lacks HDR detail replenishment and presents relatively smooth edges in invariant W . Through the iterative interaction of the proposed two modules, our framework achieves accurate texture details and effective noise suppression, leading to enhanced image quality. Dataset Release GT Resolution Dynamic Event Frame Scene Camera 78 ✗ ✗ 180 \times 240 180 \times 240 ✓ ✓ LIE 24 ✓ ✓ 346 \times 260 346 \times 260 ✗ ✗ 31 ✗ ✗ 346 \times 260 1920 \times 1280 — — SDE 30 ✓ ✓ 346 \times 260 346 \times 260 ✗ ✓ RLE (Ours) ✓ ✓ 1024 \times 768 1024 \times 768 ✓ ✓ Table 1 : Summary of existing real-world event-based low-light enhancement datasets ( Section 2.3 ). In addition, we design an optical imaging system with dual beam splitters and construct a high-resolution ( 1024\times 768 ) RLE dataset, significantly surpassing the SDE dataset ( 346\times 260 ) in resolution. This dataset includes high-quality event-image pairs spanning a wide illumination range in both indoor and outdoor environments, as well as complex dynamic scenes. It also features time-synchronized low-light images and event streams alongside corresponding normal-light images. Our comprehensive dataset overcomes the limitations in this field 78 ; 31 ; 24 ; 30 , providing a robust benchmark for event-based low-light enhancement. Overall, our contributions are summarized as follows: • We propose a new event-illumination collaborative low-light image enhancement (EIC-LIE) framework. • We formulate collaborative bidirectional interaction processes and design Event-Illumination Collaborative Interaction (EICI) that merges high dynamic range (HDR) features across various lighting conditions and complements modal-specific content for illumination and event representations. • We develop the Illumination-Aware Event Filter (IAEF) to dynamically filter event noise using bright statistics from images. • We construct a high-resolution real-world dataset (RLE) consisting of high-quality image-event pairs for evaluating the proposed framework. Experimental results demonstrate that our EIC-LIE achieves state-of-the-art performance across various benchmarks, surpassing previous studies 30 . Specifically, it shows improvements of up to 0.95dB/0.0469, 1.01dB/0.0695, and 1.24dB/0.0312 in terms of PSNR/SSIM on the RLE, SDE, and SDSD datasets, respectively. 2 Related Work 2.1 Image-based LIE Traditional non-learning low-light enhancement (LIE) methods relied on hand-crafted priors 2 ; 41 ; 12 ; 18 ; 73 , which suffer from limited adaptability and efficiency 69 . With significant advancements in deep learning 19 ; 58 ; 8 ; 71 ; 72 ; 44 ; 45 ; 49 ; 48 ; 50 ; 46 ; 51 ; 47 ; 7 ; 22 ; 29 ; 37 ; 39 , learning-based LIE methods 38 have shown substantial improvements, which can be further categorized into Retinex-based 65 ; 79 ; 69 ; 4 ; 10 ; 66 ; 42 and non-Retinex-based 40 ; 64 ; 63 ; 70 approaches. Specifically, 4 propose a one-stage Retinex-based Illumination-Guided Transformer to exploit the illumination representations to direct the computation of self-attention. 66 incorporate illumination guidance into state space models (SSMs) to enhance the Retinex-based approach. However, the crucial illumination information

Read the complete text