1 Low-Light Image Restoration Based on Retina Model using Neural Networks

2025-04-28 0 0 403.29KB 4 页 10玖币

侵权投诉

Low-Light Image Restoration Based on Retina

Model using Neural Networks

Yurui Ming, Yuanyuan Liang

Abstract—We report the possibility of using a simple neural

network for effortless restoration of low-light images inspired by

the retina model, which mimics the neurophysiological principles

and dynamics of various types of optical neurons. The proposed

neural network model saves the cost of computational overhead in

contrast with traditional signal-processing models, and generates

results comparable with complicated deep learning models from

the subjective perceptual perspective. This work shows that to

directly simulate the functionalities of retinal neurons using neural

networks not only avoids the manually seeking for the optimal

parameters, but also paves the way to build corresponding

artificial versions for certain neurobiological organizations.

Index Terms—Neural Networks, Retina Model, Low-Light

Image Restoration, Depthwise Convolution

I. INTRODUCTION

CHOLARS already discover the intricacy of the

information processing capabilities of the mammalian’s

retina [1]. It does intensive pre-processing to the optical

signals before forwarding them to higher-level visual cortices.

Recent study also reveals the bunch of subtypes of retinal

neurons, all of which are endowed with special roles to enable

the hosts to swiftly and smartly adjust to varying environments

based on perceptions [2, 3]. Although to disclose all the retina’s

secrecies folded during the evolution is still a long way to go,

by resorting to the working principles of retina, more intelligent

algorithms related to perception can be designed to generate

elegant achievements. Examples along this thread include the

retinex theory [4], photoreceptor adaptation proposition [5],

etc., which are specifically used to cater to image processing

tasks such as high-dynamic range (HDR) image tone mapping

(TM) [6-9].

The work in [10] is iconic in that it intensively considers the

working principles of various types of neurons, such as

horizontal neurons, bipolar neurons, etc., to inspire a model for

TM task. Usually, image occurred in nature can be within a

rather high dynamic range, which needs to be mapped or

compressed in a suitable range that suits the eye. Since human

eyes do the TM in some unconscious way, it is believed that the

characteristic of retina must be endowed with the capabilities in

processing dynamic scenes. In [10], the authors systematically

inspect different aspects of the retinal circuitry from the signal

processing perspective, ranging from feedbacks of the

horizontal cells to activation patterns of the bipolar cells. These

neuronal processing traits inspire the authors articulating a

corresponding computational model. By separating image into

individual channels to modulate by the algorithm and

aggregating the outcomes, this work achieves the best result

among all.

Usually, a traditional treatment of TM in digital image

processing is histogram equalization, namely, to re-adjust the

distribution of histogram in an equilibrium conforming the new

pixel value range. However, the reshuffle of pixel values tends

to cause the original image loses the colour constancy, which

means objects can have quite different or unnatural colours after

restoration. This is especially challenging for the low-light

image restoration (LIIR) task. Although compared with

traditional approaches, work in [10] is with exceptional result,

however, it is modelling the procedures by mathematical

formulas and still too algorithmic-oriented. It uses convolutions

and difference of Gaussian (DoG) functions to depict the

characteristic of optical signal processing of the corresponding

neurons, which potentially loses the biological intuition

because of the minimal chance that retina works in such a way.

Meantime, some parameters of the model depend on domain

knowledge, and it requires experience from the author to select

the most appropriate ones.

In this report, inspired by the work in [10], we re-examine

the working principles of retina and design a network to tackle

the LIIR problem. The network conforms with the processing

flow of the optical signal by different neurons and has a clear

correspondence between the neural pathway in the retina, in this

way it bears a manifest explanation towards the design

motivation. Meantime, this simple model merits the end-to-end

learning philosophy to avoid manually seeking the optimal

parameters. The experiment shows a satisfying image

restoration from the subjective perceptual perspective, and we

plan the improvement over object metric as future work.

II. COMPUTATIONAL MODEL

It is already known that for eyes, the cone photoreceptors

are dedicated to colours and the rod photoreceptors are

dedicated to illuminance. Although in low illumination

environment, activations of rod cells over cone cells render a

monochromatic or grey image, the downstream cells such as

horizontal cells (HCs), amacrine cells (ACs) are still believed

to strike for a polychromatic visual perception to benefit the

survival [11, 12]. As a consequence, [10] suggests a model

which takes two stereotypical pathways for optical signal

processing in the retina. The first is the vertical path where

signals are picked up by photoreceptors, relayed by bipolar cells

(BCs) and sent to ganglion cells (GCs). The second is the lateral

pathways where local feedbacks carry information from

horizontal cells back to photoreceptors and from amacrine cells

to horizontal cells.

In addition, cone photoreceptors mainly consist of three

types, namely, S-cones, M-cones and L-cones, each of which is

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：

1Low-LightImageRestorationBasedonRetinaModelusingNeuralNetworksYuruiMing,YuanyuanLiangAbstract—Wereportthepossibilityofusingasimpleneuralnetworkforeffortlessrestorationoflow-lightimagesinspiredbytheretinamodel,whichmimicstheneurophysiologicalprinciplesanddynamicsofvarioustypesofopticalneurons.Thepro...

展开>> 收起<<

1 Low-Light Image Restoration Based on Retina Model using Neural Networks.pdf

共4页,预览1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！

1 Low-Light Image Restoration Based on Retina Model using Neural Networks

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: