Seeing Through the Noisy Dark Towards Real-world Low-Light Image Enhancement and Denoising Jiahuan Ren12 Zhao Zhang12 Richang Hong12 Mingliang Xu3Yi Yang4 Shuicheng YAN5

2025-05-03 0 0 7.94MB 10 页 10玖币

侵权投诉

Seeing Through the Noisy Dark: Towards Real-world Low-Light Image

Enhancement and Denoising

Jiahuan Ren1,2, Zhao Zhang∗1,2, Richang Hong1,2, Mingliang Xu3,Yi Yang4, Shuicheng YAN5

1School of Computer Science and Information Engineering, Hefei University of Technology, China

2Key Laboratory of Knowledge Engineering with Big Data (Ministry of Education) & Intelligent

Interconnected Systems Laboratory of Anhui Province, Hefei University of Technology, China

3School of Information Engineering, Zhengzhou University, China

4School of Computer Science and Technology, Zhejiang University, China

5Sea AI Lab (SAIL), Singapore

(a) Input (b) RetinexNet [27](c) R2RNet [8]

(h) RUAS [17]

(d) KinD++ [35]

(g) SCI [20](k) LLFlow [26](l) Ours

(j) SNR [30]

(e) Zero-DCE++ [13]

(i) DCC-Net [37]

(g) EnlightenGAN [10]

Figure 1. LLIE comparison on a real-world image taken in noisy dark. Clearly, our RLED-Net can effectively see though the noisy dark,

remove noise and recover the details. In contrast, the enhanced images of other methods (b-k) still contain some speckle noise and blur.

Abstract

Low-light image enhancement (LLIE) aims at improving

the illumination and visibility of dark images with lighting

noise. To handle the real-world low-light images often with

heavy and complex noise, some efforts have been made for

joint LLIE and denoising, which however only achieve in-

ferior restoration performance. We attribute it to two chal-

lenges: 1) in real-world low-light images, noise is some-

what covered by low-lighting and the left noise after denois-

ing would be inevitably ampliﬁed during enhancement; 2)

conversion of raw data to sRGB would cause information

loss and also more noise, and hence prior LLIE methods

trained on raw data are unsuitable for more common sRGB

images. In this work, we propose a novel Low-light En-

hancement & Denoising Network for real-world low-light

images (RLED-Net) in the sRGB color space. In RLED-Net,

we apply a plug-and-play differentiable Latent Subspace

Reconstruction Block (LSRB) to embed the real-world im-

ages into low-rank subspaces to suppress the noise and rec-

tify the errors, such that the impact of noise during enhance-

ment can be effectively shrunk. We then present an efﬁcient

Crossed-channel & Shift-window Transformer (CST) layer

with two branches to calculate the window and channel at-

tentions to resist the degradation (e.g., speckle noise and

blur) caused by the noise in input images. Based on the

CST layers, we further present a U-structure network CST-

Net as backbone for deep feature recovery, and construct

a feature reﬁne block to reﬁne the ﬁnal features. Exten-

sive experiments on both real noisy images and public im-

age databases well verify the effectiveness of the proposed

RLED-Net for RLLIE and denoising simultaneously.

1. Introduction

Low-light image enhancement (LLIE) [4,27,35] is an

important low-level task to enhance the illumination of low-

light images to normal-light ones [7,11,24,27], beneﬁting

arXiv:2210.00545v3 [cs.CV] 15 Nov 2022

Input Reflectance Illumination Amplification

Figure 2. Image decomposition via RetinexNet. It can be observed

clearly that the illumination of a low-light image contains noise.

many visual applications for processing the low-light im-

ages with poor visibility, for instance object detection, im-

age recognition and segmentation in the dark [12]. Early

methods solve this task by designing minimal reconstruc-

tion models [11,21], which generally have limited capa-

bilities for restoring detailed information of the low-light

images. In recent years, deep neural networks have been

widely applied to LLIE tasks [15,17,28,33]. These deep

models are mostly trained on dark images with relatively

low noise, achieving good performance on recovering local

textures and global structures of the images [6,13,30,37].

Among previous deep LLIE methods, some tackle image

enhancement jointly with denoising, such as [17,27,28,35]

that are aimed at removing noise in some speciﬁc layers

while enhancing image in other components, or [6,19,32]

that are aimed at removing noise and enhance image jointly.

Although these methods achieve promising effectiveness,

they still tend to give inferior results with speckle noise

and blur when applied to real-world low-light sRGB im-

ages, as shown in Figure 1. We hold that such performance

gap mainly roots in two challenges of real-world LLIE. (1)

The noise issue in real-world LLIE is rather complex. Due

to insufﬁcient lighting or hardware limitations, the images

captured in real-world dark environments inevitably contain

varying degrees of noise, which is often with unknown dis-

tribution. Moreover, some noise is somewhat covered up by

the low-lighting, which makes it hard to be separated com-

pletely with traditional methods. Besides, the unobserved

noise in the image would be ampliﬁed and lead to speckle

noise and blur during the process of enhancement. (2) The

existing LLIE methods trained on raw data are unsuitable

for sRGB images due to the difference in the domains of

raw and sRGB data. The low-light sRGB image contain

less information compared with a raw one, and the conver-

sion of raw data to sRGB may also produce more noise.

We conduct a pilot study to investigate the separation of

noise from the low-light image [27] and provide visualiza-

tions in Figure 2. It can be observed that both reﬂectance

and illumination contain noise, and the ﬁnal enhanced re-

sult (Figure 1(b)) contains much speckle noise, which indi-

cates that the noise is hard to be separated completely by the

delicately designed layers, and that the hidden noise indeed

causes inaccurate representation.

We are then motivated to design a new Low-light En-

hancement & Denoising Network for enhancing real-world

low-light images in the sRGB color space, shorted as

RLED-Net. Instead of removing the noise in some spe-

ciﬁc layers, we propose to suppress the noise in the low-

rank/low-dimensional subspaces which can be used to char-

acterize the real-world images. In RLED-Net, a low-rank

subspace recovery block (LSRB) is used to directly embed

the noisy dark image into the low-rank subspaces to make

our network more suitable for real-world low-light image

denoising. Furthermore, to reduce the speckle noise and

blur caused by the hidden noise in the low-light environ-

ment, we propose to borrow some self-attentions to build

transformer layers for recovering the important features.

Hence we design a Crossed-channel & Shift-window Trans-

former (CST) layer with two branches. Speciﬁcally, CST

can simultaneously maintain more accurate local features

(e.g., edge/texture) and global features (e.g., color/shape),

by shift-window attention and crossed channels attention in

two parallel branches. Through the ablation studies, We ﬁnd

this structure can preserve more useful features.

We conduct extensive experiments to prove the effec-

tiveness of the proposed RLED-Net. The obtained results

demonstrate that our RLED-Net can signiﬁcantly improve

the performance when employed for real-world low-light

sRGB images enhancement and denoising. An example can

be found in Figure 1. To sum up, the main contributions of

this paper are described as follows:

• We propose a plug-and-play and differentiable low-

rank subspace recovery block (LSRB), which can sup-

press the noise in real-world low-light images effec-

tively by representing them in low-rank subspaces.

• To reduce the speckle noise and blur caused by hid-

den noise, we design a two-branch transformer (CST)

layer, which can simultaneously recover more accurate

local features and global features by self-attention.

• We develop a Low-light Enhancement & Denoising

Network (RLED-Net) for enhancing real-world low-

light images in the sRGB color space, which is a more

widely used color space in reality. SOTA performance

of RLED-Net on both well-designed and real low-light

images is veriﬁed through extensive experiments.

2. Related Work

2.1. Low-light image enhancement

Low-light image enhancement (LLIE) explores how to

enhance the illumination to improve the visibility of low-

light images [10,12,20,30,37]. Recently, some LLIE meth-

ods have been proposed to exploit the restoration of global

and local information for more natural restored results. For

example, DCC-Net [37] decomposes a color image into a

gray image and a color histogram, so as to retain the color

consistency between the enhanced image and ground truth;

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

SeeingThroughtheNoisyDark:TowardsReal-worldLow-LightImageEnhancementandDenoisingJiahuanRen1;2,ZhaoZhang1;2,RichangHong1;2,MingliangXu3,YiYang4,ShuichengYAN51SchoolofComputerScienceandInformationEngineering,HefeiUniversityofTechnology,China2KeyLaboratoryofKnowledgeEngineeringwithBigData(MinistryofEd...

展开>> 收起<<

Seeing Through the Noisy Dark Towards Real-world Low-Light Image Enhancement and Denoising Jiahuan Ren12 Zhao Zhang12 Richang Hong12 Mingliang Xu3Yi Yang4 Shuicheng YAN5.pdf

共10页,预览2页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

Seeing Through the Noisy Dark Towards Real-world Low-Light Image Enhancement and Denoising Jiahuan Ren12 Zhao Zhang12 Richang Hong12 Mingliang Xu3Yi Yang4 Shuicheng YAN5

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: