Abstract:
Building paired datasets in low-light enhancement entails significant cost and time, making such datasets precious commodities. Many researchers focus on how to enable mo...Show MoreMetadata
Abstract:
Building paired datasets in low-light enhancement entails significant cost and time, making such datasets precious commodities. Many researchers focus on how to enable models to learn more information from limited datasets. A prevalent strategy involves employing semi-supervised learning techniques to enhance model performance through additional unpaired images. However, one of the main challenges faced is the scarcity of a vast number of unpaired images from the same domain as the original low-light images. Consequently, we introduce a semi-supervised image enhancement method using pseudo-low-light images. Initially, we generate pseudo low-light images with less noise compared with the source domain image by the Signal-to-Noise Ratio prior and diffusion models. We then employ the Mean-Teacher network and the feature constraints of the pseudo-low-light images to realize low-light image enhancement. Comprehensive experimental results validate the efficacy of our approach on real-world datasets.
Published in: 2023 16th International Congress on Image and Signal Processing, BioMedical Engineering and Informatics (CISP-BMEI)
Date of Conference: 28-30 October 2023
Date Added to IEEE Xplore: 02 January 2024
ISBN Information: