EvRWKV: A Continuous Interactive RWKV Framework for Effective Event-Guided Low-Light Image Enhancement

TL;DR

EvRWKV framework enhances low-light images with continuous interaction, improving PSNR by 1.79 dB and 1.85 dB.

eess.IV 🔴 Advanced 2025-07-02 4 views
Wenjie Cai Qingguo Meng Zhenyu Wang Xingbo Dong Zhe Jin
low-light image enhancement event cameras cross-modal fusion RWKV dual-domain processing

Key Findings

Methodology

The EvRWKV framework achieves continuous cross-modal interaction through dual-domain processing, featuring the Cross-RWKV Module and the Event Image Spectral Fusion Enhancer (EISFE). The Cross-RWKV Module captures fine-grained temporal and cross-modal dependencies, while the EISFE module performs frequency-domain denoising and spatial alignment, maintaining consistency from low-level textures to high-level semantics.

Key Results

  • EvRWKV improves PSNR by 1.79 dB and 1.85 dB on the SDE and SDSD datasets, respectively, significantly outperforming image-only methods.
  • In semantic segmentation tasks, images enhanced by EvRWKV show a 35.44% improvement in mIoU.
  • Ablation studies confirm the critical role of the Cross-RWKV Module in cross-modal feature alignment.

Significance

The EvRWKV framework is significant in the field of low-light image enhancement by addressing the issues of modality heterogeneity in early fusion and feature disconnection in late fusion. It not only improves image quality but also shows significant performance gains in downstream tasks, advancing the practical application of event cameras.

Technical Contribution

EvRWKV introduces the Cross-RWKV Module and EISFE Module, achieving consistent cross-modal fusion from low-level to high-level features, overcoming limitations in feature alignment and noise suppression in traditional methods, providing a new technical pathway for low-light image enhancement.

Novelty

EvRWKV is the first framework to use the RWKV architecture for continuous cross-modal interaction, effectively combining the strengths of image and event data through dual-domain processing, addressing limitations of traditional fusion methods.

Limitations

  • In extreme low-light conditions, the sparsity of event data may lead to reduced spatial resolution.
  • Requires high-performance hardware, making it difficult to implement with standard cameras.

Future Work

Future work could explore the application of EvRWKV in other visual tasks such as video deblurring and action detection, further optimizing cross-modal interaction mechanisms.

AI Executive Summary

Event cameras hold great potential for low-light image enhancement, yet existing fusion methods face challenges with modality heterogeneity in early fusion and feature disconnection in late fusion. The EvRWKV framework achieves continuous cross-modal interaction through dual-domain processing, primarily featuring the Cross-RWKV Module and the Event Image Spectral Fusion Enhancer (EISFE). Experiments show that EvRWKV improves PSNR by 1.79 dB and 1.85 dB on the SDE and SDSD datasets, respectively, and increases mIoU by 35.44% in semantic segmentation tasks.

The EvRWKV framework introduces the Cross-RWKV Module and EISFE Module, achieving consistent cross-modal fusion from low-level to high-level features, overcoming limitations in feature alignment and noise suppression in traditional methods, providing a new technical pathway for low-light image enhancement. It not only improves image quality but also shows significant performance gains in downstream tasks, advancing the practical application of event cameras.

However, in extreme low-light conditions, the sparsity of event data may lead to reduced spatial resolution. Additionally, it requires high-performance hardware, making it difficult to implement with standard cameras. Future work could explore the application of EvRWKV in other visual tasks such as video deblurring and action detection, further optimizing cross-modal interaction mechanisms.

Deep Analysis

Background

Low-light image enhancement is a critical issue in computer vision. Traditional methods like histogram equalization and Retinex algorithms face limitations in handling complex lighting conditions. Recently, deep learning methods, particularly convolutional neural networks and Transformer architectures, have made significant progress in this field. However, these methods still suffer from information loss due to sensor noise and motion blur.

Core Problem

Existing low-light image enhancement methods struggle with modality heterogeneity in early fusion and feature disconnection in late fusion, failing to fully leverage the complementary strengths of event and image data.

Innovation

The EvRWKV framework achieves continuous cross-modal interaction through dual-domain processing, featuring innovations like the Cross-RWKV Module and EISFE Module. The Cross-RWKV Module captures fine-grained temporal and cross-modal dependencies, while the EISFE module performs frequency-domain denoising and spatial alignment, maintaining consistency from low-level textures to high-level semantics.

Methodology

  • �� Cross-RWKV Module: Utilizes the Receptance Weighted Key Value architecture for fine-grained interaction between event and image features.
  • �� EISFE Module: Performs joint denoising and alignment in frequency and spatial domains, ensuring precise alignment of structural details.
  • �� Dual-domain processing: Maintains feature consistency through continuous modality interaction.

Experiments

Experiments were conducted on the SDE and SDSD datasets, with baselines including EvLight. PSNR and mIoU were used as evaluation metrics, with various hyperparameters set for ablation studies to verify the effectiveness of each module.

Results

EvRWKV improves PSNR by 1.79 dB and 1.85 dB on the SDE and SDSD datasets, respectively, significantly outperforming image-only methods. In semantic segmentation tasks, images enhanced by EvRWKV show a 35.44% improvement in mIoU.

Applications

EvRWKV can be used for image enhancement in low-light environments, such as night surveillance and autonomous driving. It requires high-performance hardware, equipped with event cameras.

Limitations & Outlook

In extreme low-light conditions, the sparsity of event data may lead to reduced spatial resolution. Additionally, it requires high-performance hardware, making it difficult to implement with standard cameras. Future work could explore the application of EvRWKV in other visual tasks such as video deblurring and action detection, further optimizing cross-modal interaction mechanisms.

Plain Language Accessible to non-experts

Imagine you're in a dark room with a flashlight and night vision goggles. The flashlight can illuminate what's in front of you but creates lots of shadows and glare. The night vision goggles let you see outlines in the dark but lack detail. The EvRWKV framework is like a super device that combines the flashlight and night vision goggles, illuminating details while seeing outlines in the dark. By constantly adjusting the balance between the two, it provides clear images in any lighting condition.

ELI14 Explained like you're 14

Imagine playing a dark level in a game with a flashlight and night vision goggles. The flashlight lets you see what's ahead, but there are lots of shadows. The night vision goggles let you see outlines in the dark but aren't clear. EvRWKV is like a super gadget that uses both the flashlight and night vision goggles, letting you see all the details in the dark! It's like a game power-up, making it easy to beat the level!

Glossary

EvRWKV

A continuous interaction framework for low-light image enhancement, combining the strengths of event cameras and image data.

Used in the paper to achieve cross-modal interaction.

Cross-RWKV Module

Utilizes the Receptance Weighted Key Value architecture for fine-grained interaction between event and image features.

Used in the EvRWKV framework to capture temporal and cross-modal dependencies.

EISFE Module

Event Image Spectral Fusion Enhancer for frequency-domain denoising and spatial alignment.

Used in the EvRWKV framework to maintain feature consistency.

PSNR

Peak Signal-to-Noise Ratio, a metric for evaluating image enhancement effects.

Used in experimental results to quantify image quality improvement.

mIoU

Mean Intersection over Union, a metric for evaluating semantic segmentation tasks.

Used in downstream tasks to validate the practicality of enhanced images.

Open Questions Unanswered questions from this research

  • 1 How to improve the spatial resolution of event data in extreme low-light conditions?
  • 2 How to reduce dependency on high-performance hardware?

Applications

Immediate Applications

Night Surveillance

EvRWKV can enhance the clarity of night surveillance images, helping security personnel better identify potential threats.

Long-term Vision

Autonomous Driving

In autonomous driving, EvRWKV can enhance perception capabilities in low-light environments, ensuring driving safety.

Abstract

Event cameras offer significant potential for Low-light Image Enhancement (LLIE), yet existing fusion approaches are constrained by a fundamental dilemma: early fusion struggles with modality heterogeneity, while late fusion severs crucial feature correlations. To address these limitations, we propose EvRWKV, a novel framework that enables continuous cross-modal interaction through dual-domain processing, which mainly includes a Cross-RWKV Module to capture fine-grained temporal and cross-modal dependencies, and an Event Image Spectral Fusion Enhancer (EISFE) module to perform joint adaptive frequency-domain denoising and spatial-domain alignment. This continuous interaction maintains feature consistency from low-level textures to high-level semantics. Extensive experiments on the real-world SDE and SDSD datasets demonstrate that EvRWKV significantly outperforms only image-based methods by 1.79 dB and 1.85 dB in PSNR, respectively. To further validate the practical utility of our method for downstream applications, we evaluated its impact on semantic segmentation. Experiments demonstrate that images enhanced by EvRWKV lead to a significant 35.44% improvement in mIoU.

eess.IV cs.CV