Attention-Based Generative Neural Image Compression on Solar Dynamics Observatory Ali Zafariy Atefeh Khoshkhahtinaty Piyush M. Mehtaz Nasser M. Nasrabadiy Barbara J. Thompsonx

2025-04-27 0 0 1.22MB 8 页 10玖币

侵权投诉

Attention-Based Generative Neural Image

Compression on Solar Dynamics Observatory

Ali Zafari†, Atefeh Khoshkhahtinat†, Piyush M. Mehta‡, Nasser M. Nasrabadi†, Barbara J. Thompson§,

Daniel da Silva§, Michael S. F. Kirk§

†Dept. of Computer Science & Electrical Engineering, West Virginia University, WV USA

‡Dept. of Mechanical & Aerospace Engineering, West Virginia University, WV USA

§NASA Goddard Space Flight Center, MD USA

{az00004,ak00043}@mix.wvu.edu,{piyush.mehta,nasser.nasrabadi}@mail.wvu.edu

{barbara.j.thompson,daniel.e.dasilva,michael.s.kirk}@nasa.gov

Abstract—NASA’s Solar Dynamics Observatory (SDO) mission

gathers 1.4 terabytes of data each day from its geosynchronous

orbit in space. SDO data includes images of the Sun captured at

different wavelengths, with the primary scientiﬁc goal of under-

standing the dynamic processes governing the Sun. Recently, end-

to-end optimized artiﬁcial neural networks (ANN) have shown

great potential in performing image compression. ANN-based

compression schemes have outperformed conventional hand-

engineered algorithms for lossy and lossless image compression.

We have designed an ad-hoc ANN-based image compression

scheme to reduce the amount of data needed to be stored and

retrieved on space missions studying solar dynamics. In this

work, we propose an attention module to make use of both local

and non-local attention mechanisms in an adversarially trained

neural image compression network. We have also demonstrated

the superior perceptual quality of this neural image compressor.

Our proposed algorithm for compressing images downloaded

from the SDO spacecraft performs better in rate-distortion trade-

off than the popular currently-in-use image compression codecs

such as JPEG and JPEG2000. In addition we have shown that

the proposed method outperforms state-of-the art lossy transform

coding compression codec, i.e., BPG.

Index Terms—Learned lossy image compression, solar dynam-

ics observatory, generative adversarial network, attention

I. INTRODUCTION

Image compression using artiﬁcial neural networks (ANN)

has shown great potential to be applied on a wide variety

of different areas since their ﬁrst appearance [1]. In the

past couple of years, they have outperformed most of the

hand-engineered codecs such as JPEG [2] and JPEG2000

[3] in terms of rate-distortion (RD) performance [4]. One

major advantage of ANN-based compression algorithms is

that they can be developed on any ad-hoc dataset to do better

compression than general codecs [4].

Although it is believed that the ultimate trade-off in im-

age compression is between the rate and distortion, recent

studies have shown that there is a third role governing the

This research is based upon work supported by the National Aeronautics

and Space Administration (NASA), via award number 80NSSC21M0322

under the title of Adaptive and Scalable Data Compression for Deep Space

Data Transfer Applications using Deep Learning.

visual quality of compressed images known as perception [5].

Generative Adversarial Networks (GANs) are known for their

high-quality reconstructed images by enforcing the ANN to

capture the distribution of their input image. Hence, to improve

the perceptual quality of reconstructed image at the receiver,

GANs have been applied to image compression networks in

the literature [6].

Another venue of works to improve the performance of Con-

volutional Neural Networks (CNNs) is attention mechanism.

With its unprecedented inﬂuence in natural language process-

ing [7], attention has found its way in computer vision and

object detection/classiﬁcation tasks [8], [9]. We have utilized

both of these improvements in learned image compression net-

works to enhance the performance in terms of rate-distortion-

perception trade-off [5]. As shown in Figure 1, although the

attention mechanism can reach better performance compared

with other compression standards, augmenting it with a GAN

will lead to better perceptual quality.

Contributions of This Work. In this work we have inves-

tigated the application of recently successful learned image

compression methods in the ﬁeld of solar imaging. We have

shown that these neural compression schemes could easily

outperform traditional and currently-in-use image codecs. In

addition, we have proposed a curated attention module to

improve the RD tradeoff performance in state-of-the art neural

compression architectures. We have also utilized adversarial

training to encourage the decoder of our neural network

to preserve the distribution of the solar images during the

reconstruction process.

The remainder of the paper is organized as follows. Sec-

tion II reviews the neural-based compression methods and

the importance of compression on SDO mission. Section III

describes our proposed method. The experiments and ablation

studies are discussed in section IV with a conclusion at section

II. RELATED WORK

A. Neural Image Compression

Transform coding based image compression algorithms

share four main steps to compress an image [10]. First,

DOI 10.1109/ICMLA55696.2022.00035

arXiv:2210.06478v2 [eess.IV] 4 May 2023

0.367/33.27

JPEG

BPG

0.362/35.07

Original

0.361/34.10.360/35.2

Attention

JPEG2000

0.361/34.17

Attention+GAN

Fig. 1. Visual comparison of proposed compression schemes (Attention only

and GAN+Attention) to other standard codecs. Reported performance in terms

of bit-rate/distortion [bpp↓/PSNR↑]. GAN outputs are visually closer to the

original input unless their lower performance in terms of PSNR. Best viewed

on screen.

encoding the images from their input space (e.g., RGB) to an

uncorrelated space. Second, quantizing to discard less signiﬁ-

cant information from the data in its uncorrelated domain. At

the third step, an entropy coding will be utilized to losslessly

encode the quantized samples into a stream of ones and zeros.

This bitstream will be the compressed image. Final step occurs

at the receiving end (or at reconstructing step), which is

responsible to decode the quantized values to the original

space of the input image. The ﬁrst and most widely used

architecture to mimic this scenario in deep neural networks, is

the convolutional autoencoder, which has shown its superiority

in the literature [11]. Both the encoding and decoding part

of the traditional transform coding, could be imitated by an

autoencoder [12].

End-to-end optimizing of the neural networks are capable

of handling various tasks [13], [14], [15], [16] if the learning

objective chosen to be differentiable. In an end-to-end opti-

misation of an autoencoder, problems arise when we want

to do quantization on its bottleneck. It is worth mentioning

that quantization is the essential part of compression. Merely

doing dimensionality reduction cannot necessarily result in

discarding the redundant information, which is necessary to

attain high compression ratios [17].

ANNs are optimized using gradient descent algorithms

which update the parameters of the network by back-

propagating the gradients of the loss function. Thus, all the

operations performed in it must be differentiable. As a result,

we need to approximate hard discrete quantization with a soft

continuous operation. To do so, several approaches have been

proposed in the literature. [1] used recurrent neural networks

to directly binarize the latent code stochastically, while [17]

used an approach similar to straight through estimator [18]

by back-propagating the gradients of identity function and

rounding to the nearest integer in the forward pass. By this

continuous approximation, the network parameters can be

successfully learned with backpropagating the gradient of the

loss function. The most widely used approach is proposed by

[19], inherited from [20], they showed that adding independent

and identically distributed uniform noise in the range of scalar

quantization can be interpreted equivalently as doing scalar

quantization on the bottleneck. By doing so, we can optimize

the differential entropy of the continuous approximation as

a variational upper bound [21] to reduce the entropy of

the bottleneck. Low entropy messages are compressed more

efﬁciently into bitstreams [22], [23].

In classical image compression schemes, to get the best

out of the quantization process, the ﬁrst step was to apply

an invertible linear transform and translate the image into

decorrelated coefﬁcients using Discrete Cosine Transform

(DCT). By doing so, scalar quantization could reach a rea-

sonable performance close to vector quantization [10]. The

application of vector quantization in ANN-based compression

has been investigated by [24], with the cost of complicated

training procedure. On the other hand, it has been shown [11],

[12] that a joint-optimized learned nonlinear transform, i.e.,

neural network, followed by scalar quantization can ideally

approximate a parametric form of vector quantization.

Replacing the actual quantization with uniform noise ap-

proximation in the bottleneck of a vanilla autoencoder during

training of the network [25], will transform it to a Variational

Aueoencoder (VAE) [26]. The only difference is on the chosen

prior. In autoencoder-based image compression, the Gaussian

prior of the VAE is replaced with unit uniform distribution

centered on integer numbers.

B. Solar Dynamics Observatory (SDO) Mission

1) Image Compression on SDO Data:Advances in sensor

technology and an increasing desire for a deeper understanding

of the space environment (Sun to Earth and beyond) have led

to an explosion of data volume in recent years (unprecedented

spatial and/or temporal resolution as well as multi-spectral

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：

Attention-BasedGenerativeNeuralImageCompressiononSolarDynamicsObservatoryAliZafariy,AtefehKhoshkhahtinaty,PiyushM.Mehtaz,NasserM.Nasrabadiy,BarbaraJ.Thompsonx,DanieldaSilvax,MichaelS.F.KirkxyDept.ofComputerScience&ElectricalEngineering,WestVirginiaUniversity,WVUSAzDept.ofMechanical&AerospaceEngineer...

展开>> 收起<<

Attention-Based Generative Neural Image Compression on Solar Dynamics Observatory Ali Zafariy Atefeh Khoshkhahtinaty Piyush M. Mehtaz Nasser M. Nasrabadiy Barbara J. Thompsonx.pdf

共8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！

Attention-Based Generative Neural Image Compression on Solar Dynamics Observatory Ali Zafariy Atefeh Khoshkhahtinaty Piyush M. Mehtaz Nasser M. Nasrabadiy Barbara J. Thompsonx

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: