Unsupervised Boosted Fusion Network for Single Low-Light Image Enhancement

Unsupervised methods are gradually becoming a research hotspot in low-light image enhancement due to their lack of paired training data and higher quality visual enhancement effects, but existing unsupervised low-light image enhancement methods generate excess noise and rely on contrast exposure sta...

Full description

Saved in:
Bibliographic Details
Main Authors: Jianfeng Zhang, Hengxuan Li, Zhanqiang Huo
Format: Article
Language:English
Published: IEEE 2024-01-01
Series:IEEE Access
Subjects:
Online Access:https://ieeexplore.ieee.org/document/10758638/
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1850066072368054272
author Jianfeng Zhang
Hengxuan Li
Zhanqiang Huo
author_facet Jianfeng Zhang
Hengxuan Li
Zhanqiang Huo
author_sort Jianfeng Zhang
collection DOAJ
description Unsupervised methods are gradually becoming a research hotspot in low-light image enhancement due to their lack of paired training data and higher quality visual enhancement effects, but existing unsupervised low-light image enhancement methods generate excess noise and rely on contrast exposure stacks. To address this issue, we propose an unsupervised boosted fusion network for single low-light image enhancement. In the network, we generate four images with different illuminations and texture detail information from the input low-light image by applying inverting, fixed-value gamma correction, adaptive gamma correction, and color-preserving adaptive histogram equalization in the preprocessing stage. Then, we introduce DenseNet and design dedicated fusion weights to progressively fuse the four images generated in previous stage, which can boost the image contrast and brightness while suppressing overexposure phenomena in the bright regions. Finally, a self-supervised denoising network based on UNet++ is designed for suppressing the amplified noise while maintaining brightness, contrast, and texture details. The visual evaluation and quantitative experiments demonstrate that our proposed method outperforms the current state-of-the-art unsupervised low-light image enhancement methods. Especially, the enhancement effect in some scenes is superior to the supervised low-light image enhancement networks. The UBF-Net code is available at.<uri>https://github.com/huozhanqiang/UBF-Net</uri>.
format Article
id doaj-art-8ead61e711464d8c8757b2223ebbaa6c
institution DOAJ
issn 2169-3536
language English
publishDate 2024-01-01
publisher IEEE
record_format Article
series IEEE Access
spelling doaj-art-8ead61e711464d8c8757b2223ebbaa6c2025-08-20T02:48:51ZengIEEEIEEE Access2169-35362024-01-011217925217926410.1109/ACCESS.2024.350265410758638Unsupervised Boosted Fusion Network for Single Low-Light Image EnhancementJianfeng Zhang0https://orcid.org/0009-0007-1408-4064Hengxuan Li1https://orcid.org/0009-0007-2193-5445Zhanqiang Huo2https://orcid.org/0000-0001-9243-5009College of Information Engineering, Henan Vocational College of Agriculture, Zhengzhou, ChinaElectronic Control Department, Zhengzhou Hengda Intelligent Control Technology Company Ltd., Zhengzhou, ChinaSchool of Software, Henan Polytechnic University, Jiaozuo, ChinaUnsupervised methods are gradually becoming a research hotspot in low-light image enhancement due to their lack of paired training data and higher quality visual enhancement effects, but existing unsupervised low-light image enhancement methods generate excess noise and rely on contrast exposure stacks. To address this issue, we propose an unsupervised boosted fusion network for single low-light image enhancement. In the network, we generate four images with different illuminations and texture detail information from the input low-light image by applying inverting, fixed-value gamma correction, adaptive gamma correction, and color-preserving adaptive histogram equalization in the preprocessing stage. Then, we introduce DenseNet and design dedicated fusion weights to progressively fuse the four images generated in previous stage, which can boost the image contrast and brightness while suppressing overexposure phenomena in the bright regions. Finally, a self-supervised denoising network based on UNet++ is designed for suppressing the amplified noise while maintaining brightness, contrast, and texture details. The visual evaluation and quantitative experiments demonstrate that our proposed method outperforms the current state-of-the-art unsupervised low-light image enhancement methods. Especially, the enhancement effect in some scenes is superior to the supervised low-light image enhancement networks. The UBF-Net code is available at.<uri>https://github.com/huozhanqiang/UBF-Net</uri>.https://ieeexplore.ieee.org/document/10758638/Low-light image enhancementunsupervised learningimage fusionnoise suppression
spellingShingle Jianfeng Zhang
Hengxuan Li
Zhanqiang Huo
Unsupervised Boosted Fusion Network for Single Low-Light Image Enhancement
IEEE Access
Low-light image enhancement
unsupervised learning
image fusion
noise suppression
title Unsupervised Boosted Fusion Network for Single Low-Light Image Enhancement
title_full Unsupervised Boosted Fusion Network for Single Low-Light Image Enhancement
title_fullStr Unsupervised Boosted Fusion Network for Single Low-Light Image Enhancement
title_full_unstemmed Unsupervised Boosted Fusion Network for Single Low-Light Image Enhancement
title_short Unsupervised Boosted Fusion Network for Single Low-Light Image Enhancement
title_sort unsupervised boosted fusion network for single low light image enhancement
topic Low-light image enhancement
unsupervised learning
image fusion
noise suppression
url https://ieeexplore.ieee.org/document/10758638/
work_keys_str_mv AT jianfengzhang unsupervisedboostedfusionnetworkforsinglelowlightimageenhancement
AT hengxuanli unsupervisedboostedfusionnetworkforsinglelowlightimageenhancement
AT zhanqianghuo unsupervisedboostedfusionnetworkforsinglelowlightimageenhancement