핵심 개념
Zero-LED proposes a novel zero-reference lighting estimation diffusion model for low-light image enhancement, bridging the gap between low-light and normal-light domains without paired training data.
초록
Low-light image enhancement aims to improve image quality under challenging lighting conditions.
Traditional and deep learning-based approaches have been developed for low-light image enhancement.
Diffusion models have shown promising results but face challenges in unsupervised training.
Zero-LED introduces a bidirectional unsupervised diffusion training approach for effective low-light image enhancement.
Appearance Reconstruction Module guides image restoration through semantic and frequency domain-based reconstruction.
Extensive experiments demonstrate the superiority of Zero-LED over state-of-the-art methods.
통계
"Extensive experiments demonstrate the superiority of our approach over other state-of-the-art methods and more significant generalization capabilities."
"Our method achieves quantitative performance close to the state-of-the-art on several metrics compared to all compared methods."
인용구
"We propose a bidirectional optimised unsupervised training method that effectively implements a low-light image enhancement diffusion model without relying on reference images, thus reducing the dependence on paired training data."
"We design a semantic and frequency domain-based appearance reconstruction module. It utilizes different modalities and multiple frequency domain spaces to constrain the stochastic nature of the diffusion inference process and efficiently reconstructs images for better perceptual results."