Dual-Exposure Imaging with Events

Mingyuan Lin; Hongyi Liu; Chu He; Wen Yang; Gui-Song Xia; Lei Yu

cs.CVApr 11, 2026

Dual-Exposure Imaging with Events

Authors: Mingyuan Lin, Hongyi Liu, Chu He, Wen Yang, Gui-Song Xia, Lei Yu

Organizations: School of Electronic Information, Wuhan University, Wuhan 430072, China · School of Artificial Intelligence, Wuhan University, Wuhan 430072, China

Abstract

By combining complementary benefits of short- and long-exposure images, Dual-Exposure Imaging (DEI) enhances image quality in low-light scenarios. However, existing DEI approaches inevitably suffer from producing artifacts due to spatial displacement from scene motion and image feature discrepancies from different exposure times. To tackle this problem, we propose a novel Event-based DEI (E-DEI) algorithm, which reconstructs high-quality images from dual-exposure image pairs and events, leveraging high temporal resolution of event cameras to provide accurate inter-/intra-frame dynamic information. Specifically, we decompose this complex task into an integration of two sub-tasks, i.e., event-based motion deblurring and low-light image enhancement tasks, which guides us to design E-DEI network as a dual-path parallel feature propagation architecture. We propose a Dual-path Feature Alignment and Fusion (DFAF) module to effectively align and fuse features extracted from dual-exposure images with assistance of events. Furthermore, we build a real-world Dataset containing Paired low-/normal-light Images and Events (PIED). Experiments on multiple datasets show the superiority of our method. The code and dataset are available at github.

Figures & tables

Fig. 1: (a) Illustration of the proposed Event-based Dual-Exposure Imaging (E-DEI) strategy. (b-c) The short- and long-exposure image pairs are continuously captured under the dynamic scenes. The corresponding image reconstruction results via (d) the state-of-the-art Frame-based DEI (F-DEI) method and (e) our E-DEI. Due to the missing inter/intra-frame motion information, F-DEI suffers from texture blur and artifacts, while E-DEI can successfully generate the high-quality visual images.

Panel of Figure 2 — Fig. 2: Comparison between (a) existing Frame-based Dual-Exposure Imaging (F-DEI) and (b) our Event-based Dual-Exposure Imaging (E-DEI). The key idea of our E-DEI is to introduce the high temporal resolution event data to complement the missing intra- and inter-frame motion information. As a result, our algorithm produces sharp and clear images without artifacts.

Panel of Figure 3 — Fig. 3: (a) The overall architecture of the proposed network of Event-based Dual-Exposure Imaging (E-DEI), which is a dual-path feature propagation and supervision framework. (b) The proposed Dual-path Feature Alignment and Fusion (DFAF) module, where features of the Deblurring path are spatially aligned to the short-exposure target time with features of the Enhancement path as the temporal reference, and then we integrate features of the two paths for the high-quality reconstruction.

Fig. 4: Details of the proposed Cross-Guided Fusion (CGF) module, which is designed to fuse the results and features of the Deblurring and Enhancement paths for the final result.

Panel of Figure 5 — Fig. 5: (a) The illustration of a beam splitter-based hybrid camera system for building the PIED dataset. (b) PIED contains the normal- and low-light image pairs (upper row) and the corresponding events (bottom row) with three diverse motion patterns.

Dataset	Image		Resolution	FPS	Motion pattern
Dataset	LL	NL	Resolution	FPS	Motion pattern
SDE [ 21 ]	✓	✓	346 $\times$ 260	$<$ 30	Camera motion
EVFI-LL [ 20 ]	✓	✗	784 $\times$ 611	64	Moving object
RLED [ 19 ]	✗	✓	1120 $\times$ 660	10,25	Camera+Object
Ours	✓	✓	1280 $\times$ 720	72	Camera+Object

TABLE I: Comparisons of our dataset with the existing event-based low-light datasets. LL/NL means low-/normal-light images

Fig. 6: Statistical properties of the E-DEI dataset. Distributions of (a) Motion magnitude, (b) Illumination, (c) Texture richness, and (d) Event firing rate.

Method		Event	REDS-LL		PIED
Method		Event	PSNR $\uparrow$	SSIM $\uparrow$	PSNR $\uparrow$	SSIM $\uparrow$
Low-light image enhancement	Uretinex-Net [ 23 ]	✗	25.621	0.9020	25.829	0.8613
	EMNet [ 70 ]	✗	30.317	0.9253	26.776	0.8982
	EvLowLight [ 18 ]	✓	25.519	0.9217	26.808	0.8666
	EvLight [ 21 ]	✓	31.211	0.9375	28.203	0.9102
Motion deblurring	MSSNet [ 58 ]	✗	30.181	0.8883	28.890	0.8538

TABLE II: Quantitative comparisons of the proposed E-DEI to the state-of-the-art methods on the REDS-LL and PIED datasets. The best performance is highlighted in bold and second best is underlined .

Panel of Figure 9 — Fig. 7: Qualitative comparisons with state-of-the-art methods on the REDS-LL (upper low) and PIED (lower row) datasets. For a better view, details are zoomed in and the brightness of short-exposure images is rescaled.

Ex.	Event		PSNR $\uparrow$	SSIM $\uparrow$
Ex.	Deblurring	Enhancement	PSNR $\uparrow$	SSIM $\uparrow$
1	✗	✗	34.522	0.9471
2	✓	✗	37.756	0.9684
3	✗	✓	36.281	0.9601
4	✓	✓	37.895	0.9695

TABLE III: Ablation study on the feeding strategies of the event streams.

Architecture	$\hat{L}_{s}$		$\hat{L}$
Architecture	PSNR $\uparrow$	SSIM $\uparrow$	PSNR $\uparrow$	SSIM $\uparrow$
Serial pipeline	35.984	0.9581	36.702	0.9658
Parallel design (Ours)	37.895	0.9695	37.904	0.9695

TABLE IV: Quantitative comparisons of the proposed E-DEI, which is a parallel architecture, to the serial pipeline design.

Ex.	Backbone		DFAF module		PSNR $\uparrow$	SSIM $\uparrow$
Ex.	Enc.	Dec.	DA	CAF	PSNR $\uparrow$	SSIM $\uparrow$
1	SP	SP	✗	✗	37.376	0.9657
2	DP	SP	✓	✓	37.637	0.9673
3	DP	DP	✗	✗	36.199	0.9591
4	DP	DP	✓	✗	35.876	0.9645
5	DP	DP	✗	✓	37.624	0.9674

TABLE V: Ablation study of the dual-path backbone and modules of our network on the REDS-LL dataset. Enc. and Dec. indicate the architecture of the encoder and decoder. DP and SP mean the dual- and single-path backbone.

Panel of Figure 13 — Fig. 8: Qualitative ablations of the backbone and modules of our network and their absolute differences to the ground-truth clear image. Enc. and Dec. indicate the architecture of the encoder and decoder.

Panel of Figure 14 — Fig. 9: Visualization of intermediate features from the DFAF module of our E-DEI. (a) and (d) the input dual-exposure image pairs. (b-c) $F_{l}$ and $F^{\prime}_{l}$ are the features before and after the DA block in the Delurring path. (e-f) $F_{s}$ and $F^{\prime}_{s}$ are the features before and after the CAF block in the Enhancement path.

Fig. 10: Evaluation of our E-DEI supervised by Eq. 20 with the different values of the balancing parameter

\lambda_{3}

in the first stage training.

Fig. 11: (a) Sensitivity analysis of the E-DEI network to the temporal misalignment between cross-modal image pairs and event streams. (b) Generalization of the E-DEI network to the unseen exposure-time ratios.

Method	Event	Param	Runtime	FLOPs
Method	Event	(M) $\downarrow$	( $ms$ ) $\downarrow$	(G) $\downarrow$
Uretinex-Net [ 23 ]	✗	0.34	123.7	1601.25
EMNet [ 70 ]	✗	12.52	774.4	1036.33
EvLowLight [ 18 ]	✓	15.03	766.9	4091.00
EvLight [ 21 ]	✓	22.73	674.6	2731.53
MSSNet [ 58 ]	✗	15.59	5.4	2040.46

TABLE VI: Quantitative comparisons of the network parameter and inference efficiency on the REDS-LL datasets. The items of runtime and FLOPs are evaluated with the input images and events at the resolution of

1280\times 720

Panel of Figure 18 — Fig. 12: Failure cases under non-ideal scenes, e.g. , extremely dim environment (upper), blocked and discarded events (middle), and coupled degradations (lower).

Explore similar work

Event-Illumination Collaborative Low-light Image Enhancement with a High-resolution Real-world Dataset
May 21, 2026Senyan Xu, Zhijing Sun, Kean Liu +5Event-Based Vision Low-Light Image Enhancement
University of Science and Technology of China
NEC-Diff: Noise-Robust Event-RAW Complementary Diffusion for Seeing Motion in Extreme Darkness
Mar 20, 2026Haoyue Liu, Jinghan Xu, Luxin Feng +4Low-Light Image Enhancement Image Denoising
National Key Lab of Multispectral Information Intelligent Processing Technology School of Artificial Intelligence and Automation, Huazhong University of Science and Technology · School of Computing, National University of Singapore
BiEvLight: Bi-level Learning of Task-Aware Event Refinement for Low-Light Image Enhancement
Mar 5, 2026Zishu Yao, Xiang-Xiang Su, Shengning Zhou +3Low-Light Image Enhancement Event-Based Vision
College of Computer and Data Science, Fuzhou University, Fuzhou 350108, China · College of Computer Science and Technology, Shandong Technology and Business University, Yantai 264003, China

May 21, 2026J/K move · Enter open · S save

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

Senyan Xu, Zhijing Sun, Kean Liu, Xin Lu+4

University of Science and Technology of China

Event-Based Vision Low-Light Image Enhancement

Open paper AskPDF arXiv

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.

Abstract

Explore similar work

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

NEC-Diff: Noise-Robust Event-RAW Complementary Diffusion for Seeing Motion in Extreme Darkness

BiEvLight: Bi-level Learning of Task-Aware Event Refinement for Low-Light Image Enhancement