ed

SwinTransEntropy: A swin-transformer based entropy

iew
model for image compression
Yujin Zhanga , Jixiang Chenga,∗, Zhidan Lia
a Schoolof Electrical Engineering and Information,
Southwest Petroleum University, Chengdu, P.R. China 610500

ev
Abstract

r
With the development of deep learning technology, the application of deep learn-
ing to image compression has received extensive attention. Although the varia-

er
tional autoencode based methods have surpassed traditional compression meth-
ods in terms of rate-distortion performance, they still possess the inherent de-
fect of convolutionary neural network, i.e., ignoring the correlation between lo-
pe
cal features and global semantic information. Inspired by the recent research of
Swin-Transformer, which leverages local perceptual attention and global percep-
tual attention to capture the correlation between spatially adjacent elements and
global semantic information, this paper proposes a Swin-Transformer based en-
tropy model for image compression, called SwimTransEntropy. This model ap-
plies a Swim-Ttransformer based codec, decouples the mean and variance param-
ot

eters of the entropy model and uses fusion context autoregressive to improve the
accuracy and efficiency of entropy coding. In addition, an asymmetric decoder
that blends traditional convolutionary neural network and Swin-Transformer is
tn

applied to enhance the capacity of recovering texture information. The results of

comprehensive experiments indicate that our method outperformes to some of the
state-of-the-art methods, and achieves high perceptual fidelity on reconstructing
images.
rin

Keywords: Image Compression, Deep learning, Swin-Transformer, Entropy

Model
ep

∗ Corresponding author.
Email addresses: 849190069@[Link] (Yujin Zhang), chengjixiang0106@[Link]
(Jixiang Cheng), dan.807@[Link] (Zhidan Li)
Pr

Preprint submitted to Digital Signal Processing April 19, 2023

This preprint research paper has not been peer reviewed. Electronic copy available at: [Link]
 ed
1. Introduction

iew
With vigorous development of information technology, images have become
an important information carrier. A large amount of image data will cause huge
transmission and storage pressure. To effectively solve this problem, image com-
pression methods are widely used. Image compression methods can be divided
into lossless compression and lossy compression. In lossless compression, the re-
construction result does not change any information of the original image, there-

ev
fore it is often used in scenes that require finely describing image details. On the
contrary, lossy image compression does not store the original image data, but the
lossy image information. On the premise of not losing image reconstruction abil-
ity, the redundant information of the image can be reduced, and the image can be

r
better restored without losing effective information. Although lossy compression
will cause some unimportant information lost and cannot fully restore the original

er
image information, it achieves a higher compression ratio and requires less stor-
age space, and is usually applied to compress natural images. Therefore, lossy
image compression is one of the most important basic supporting technologies in
pe
computer vision.
The traditional lossy image compression methods such as JPEG [1], JPEG2000
[2] and BPG [3] use well-designed basic modules, including transformation, quan-
tization and coding, to realize image compression. The most common method
JPEG is a hybrid method that uses discrete cosine transform [4] to separate high-
frequency information from low-frequency information, allocates bits according
ot

to the importance of the signal to reduce information redundancy, and uses statis-
tical coding to improve compression performance. To eliminate information re-
dundancy, the statistical coding method often uses statistical characteristics of the
tn

information source or establishes a statistical model for the information source,

and the commonly used statistical coding methods include Huffman coding, RLE
coding, arithmetic coding, etc. Since the modules are artificially designed, it is
difficult for traditional lossy image compression methods to achieve satisfactory
rin

results under various evaluation indicators. In recent years, with the rapid de-
velopment of deep learning, deep learning based image compression methods,
especially the lossy compression methods, have gradually surpassed traditional
methods and attracted great attention from both research and engineering fields.
ep

Currently, most deep learning based image compression methods are end-to-
end methods using variational autoencoder (VAE) [5]. As research has progressed,
significant progress has been made in various technical components of the archi-
tecture. Based on VAE, the image is converted into compressible latent features
Pr

This preprint research paper has not been peer reviewed. Electronic copy available at: [Link]
 ed
and then the features are quantized. Through hyperprior module [6], autoregres-

iew
sive module [7], and entropy module [6], the statistical distribution of potential
features are construct and the entropy model is estimated. The potential features
are compressed into bit stream by arithmetic coding [8]. For decoding, arith-
metic coding decompresses and dequantizes the bit stream, and then reconstructs
the image by decoder. As the quantizers are non-differentiable, they can be ap-
proximated via binarization [9], additive uniform noise [10] or random rounding

ev
[11] for directional propagation during the training phase. Although the great
progress has been made, a key issue with VAE based model is that the convolu-
tion only considers the semantic information in the artificially designed receptive
field, rather than the global semantic information. The high-level semantic in-

r
formation is obtained by crudely fusing the low-level feature information, which
ignores the correlation between the part and the whole.
Inspired by the success of the attention mechanism in natural language pro-

er
cessing [12] and computer vision [13, 14], many researchers use attention mecha-
nism to obtain the correlation weight between each feature and all other features to
guide latent feature adaptive processing, that is, allocating more bits to challeng-
pe
ing areas such as edges and textures for better rate-distortion (RD) performance.
However, attention mechanism still does not change the inherent ignorance of
the correlation between local features and global semantic information in con-
volutionary neural network(CNN). The multi-head attention mechanism in trans-
former [15] can enhance the ability of focusing on multiple places and effectively
ot

alleviate the above problem. Recent studies [16–18] show that transformer can
achieve better performance compared than convolutional neural network in vision
tasks. The Vision transformer [17] starts from the low-level features for global
association. The global correlation of the input is modelled by a multi-head at-
tn

tention mechanism, and the intermediate features are continuously linearly trans-
formed by a fully connected layer. Swin-Transformer [19] is a layered transformer
specifically designed for visual tasks with the two main features, the sliding win-
dow and the hierarchical structure. On the one hand, sliding windows compute
rin

self-attention in locally non-overlapping windows and allow connections across

windows. Through serial window self-attention operation (W-MSA) and sliding
window self-attention operation (SW-MSA), Swin-Transformer can focus on ad-
jacent elements and enhance local information extraction capabilities, and obtain
ep

near-global attention capabilities. On the other hand, the hierarchical structure

enables the model to gradually expand the receptive field.
Image compression requires not only global semantic information but also
local semantic information, and spatially adjacent elements have stronger cor-
Pr

This preprint research paper has not been peer reviewed. Electronic copy available at: [Link]
 ed
relations. On this basis, this paper proposes an Swin-Transformer based entropy

iew
model with a high compression ratio for image compression, called SwinTransEn-
troy. First, the model uses a Swin-Transformer based coder to capture the corre-
lation between spatially adjacent elements and global semantic information. Sec-
ond, the model decouples the mean and variance parameters of the entropy model
and uses fusion context autoregressive model to improve parameter estimation ac-
curacy. Finally, an asymmetric decoder combining CNN and Swin-Transformer

ev
is applied to reconstruct the image. Extensive experimental results show that our
method outperforms state-of-the-art (SOTA) image compression methods on fun-
damental metrics. The novelties of this work are as follows:

1. We design encoding and decoding modules based on the Swin-Transformer

r
by exploiting its local attention mechanism, global attention mechanism
and hierarchical structure to gradually expand receptive field to obtain more

er
compact latent features. To the best of our knowledges, this is the first work
of applying the Swin-Transformer to image compression task.
2. We decouple the mean and variance of the entropy model, and design the
pe
fusion context autoregressive model to improve the accuracy of the entropy
model. The fusion context autoregressive model expands the autoregressive
model into three mask convolutions with kernel sizes of 3*3, 5*5, and 7*7,
and uses the add method for data fusion to achieve more accurate entropy
estimation.
ot

3. Finally, we design an asymmetric decoder. Combining the traditional CNN

and Swin-Transformer increases the number of decoder model layers, which
allows the encoder to focus more on extracting features, and make the de-
coder to have enough capacity to learn reconstruction ability.
tn

4. Experimental results show that our method is superior to several SOTA

methods, and achieves high perceptual fidelity on reconstructing images.

The remainder of this paper is organized as follows. Section 2 reviews some

rin

related work and gives the motivation of the work. Section 3 details the proposed
method. Section 4 presents the experimental results on several public datasets.
Conclusions are made in Section 5.
ep

2. Related works and motivation

In recent years, lossy image compression methods based on deep learning
[6, 7, 10, 20–24] have flourished. Among them, the most widely researched are
Pr

This preprint research paper has not been peer reviewed. Electronic copy available at: [Link]
 ed
end-to-end methods based on VAE [10]. The schematic diagram of basic lossy

iew
image compression methods based on VAE is shown in Figure 1. The model
mainly includes codec module, hyper prior module, context autoregressive mod-
ule and entropy model. The codec consists of an encoder and a decoder that are
implemented by deep neural networks. The original image is encoded into latent
features which are then quantized and encoded into a bit stream. During encod-
ing and decoding, the adopted probability distribution function and cumulative

ev
distribution function depend on the estimation of the entropy model. The hyper
prior module consists of a hyperencoder and a hyperdecoder. The autoregressive
module is a context model. The hyper prior and autoregressive modules are com-
bined and applied to the entropy model in order to improve the accuracy of the

r
entropy estimation. In this section, we briefly reviewed some related work on
each module.

Input
Image
Encoder
er Hyper
Encoder
pe
Quantization Quantization

Arithmetic Encoder Arithmetic Encoder

Bit Stream (μͫσ) Context Model Bit Stream

ot

Arithmetic Decoder
Arithmetic Decoder

Output Hyper
Decoder
Image Decoder
tn

Figure 1: Schematic diagram of general VAE based lossy image compression.

rin

2.1. Codec module

The codec module consists of an encoder and a decoder. The encoder is to
transform an image to into feature maps and the decoder is to reconstruct the im-
ep

age. The performance of image compression methods heavily depends on how

the network structure is designed to produce more compact latent features. Balle
[20] proposed a general transform coding framework using generalized division
normalization module [25] for local gain during encoding and decoding, which is
Pr

This preprint research paper has not been peer reviewed. Electronic copy available at: [Link]
 ed
to improve the ability of capturing statistical characteristics of the image. Hu [26]

iew
proposed a deeper analytical transformation method for latent features to perform
more complex synthetic analytical transformations on image features while pre-
serving information. Liu [27] applied a 3D-CNN to capture redundant information
within and between channels. Ayzik [28] introduced side information to assist the
decoder side for fast image decoding. Mentzer [29] used a generative adversarial
network as a distortion metric in the training phase to guide the decoder to gener-

ev
ate more realistic texture [Link] work [30] proposed a gain unit to assign
a small weight to the channels that do not contribute prominently to the recon-
structed image in order to eliminate channel redundancy. Cheng [21] proposed
to use attention mechanism to increase the attention to complex regions for en-

r
coding and decoding modules. Later, Gao [31] proposed an attention guided two
layer image compression that uses compressive sensing to compress information.
Qian [32] proposed a dual-attention fusion mechanism to segment high-frequency

er
and low-frequency information of the input image and perform channel and space
fusion for their potential features.
pe
2.2. Hyperprior Module
The latent features extracted by the encoder have spatial redundant informa-
tion. The fully decomposed entropy model cannot capture such redundant infor-
mation while a hyperprior module can. Balle [6] proposed an entropy estimation
model with hyperprior for the first time, which introduces side information to
ot

capture hidden information represented by hidden layers and helps to generate

entropy model parameters, thereby alleviating the mismatch between the entropy
model estimation and the actual marginal distribution of hidden layers. Hu [26]
proposed a multi-layer hyperprior model to reduce spatial redundancy. Cui [22]
tn

introduced a gain unit into the hyperprior module to improve the ability of extract-
ing uneven channel redundancy in latent features, so that the network can allocate
more bits to the channels that affect the quality of reconstruction.
rin

2.3. Context autoregressive module

In deep learning based image compression, hyper prior models and context
autoregressive models are complementary on compression performance. On the
one hand, the hyperprior can learn to reserve necessary information and reduce
ep

uncertainty in the context autoregressive module. On the other hand, the con-
text autoregressive module can help the hyperprior module to introduce more side
information [6] without compromising encoding rate, because the prediction is
based only on the causal context, i.e., the already decoded latent features. Inspired
Pr

This preprint research paper has not been peer reviewed. Electronic copy available at: [Link]
 ed
by the success of autoregressive priors in probabilistic generative models, Minnen

iew
[7] first introduced the contextual autoregressive model, using the neighbourhood
correlation of predicted pixels to improve the performance of the entropy model.
Guo [33] proposed a 3D autoregressive module, which divides the discrete vari-
ables into two groups according to the channel dimension. To some extent, this
module can exploit the redundancy of two sets of potential features in the chan-
nel dimension to achieve more effective bit rate prediction. Cui [22] proposed

ev
a parallel context autoregressive module to extend mask convolution in order to
fully exploit the correlation between predicted pixels. However, since there is a
causal relationship in the original unidirectional context autoregressive module,
serial operations are required during the decompressing stage, which takes much

r
longer time. He [34] proposed a parallelized checkerboard context model and
modified the decoding order, which accelerates the decoding without affecting
the performance. Also, the work [33] extended the context autoregressive model

er
to a bidirectional parallel context autoregressive model to speed up the decoding
process.
pe
2.4. Entropy Module
Accurate entropy estimation is one of the key factors in achieving good image
compression performance. In end-to-end compression methods, the actual dis-
tribution of latent features is unknown and is estimated by the entropy module.
When the estimation is close to the actual distribution, the cross-entropy between
ot

two distributions is small, and when the two distributions are equal, the cross en-
tropy reaches the minimum and the theoretical minimum average code length is
achieved. Balle [6] proposed an SGM-based entropy model that uses a hyper prior
network to model each potential feature variable into a zero mean Gaussian distri-
tn

bution. Later, Minnen [7] further improved the entropy model by adding a context
autoregressive module and considering mean prediction. Minnen [21] proposed
an entropy model based on a discretised Gaussian mixture model (GMM), which
uses the discrete Gaussian mixture likelihood to estimate the distribution of la-
rin

tent features, thereby eliminating information redundancy and achieving accurate

entropy estimation. The above entropy models are all symmetric Gaussian distri-
butions, but the degrees of freedom are not sufficient and large estimation errors
can occur for natural images that do not satisfy the symmetric Gaussian distri-
ep

bution. Therefore, Cui [22] applied an asymmetric Gaussian entropy model to

reduce estimation errors.
Pr

This preprint research paper has not been peer reviewed. Electronic copy available at: [Link]
 ed
2.5. Motivation

iew
Inspired by the successful application of Transformer in the field of computer
vision, Qian et, al. [32] first proposed a Transformer-based image compression
model called Entroformer, which uses the mutil-head attention in Transformer
to extract global redundant information. However, the Transformer has not been
fully explored in the field of image compression. On the one hand, since multi-
head attention is computed over the entire feature map, the computational com-

ev
plexity is proportional to the square of the image size. On the other hand, the orig-
inal transformer does not have multi-scale prediction capabilities, which limits its
application. For image compression tasks, more compact latent features need to be
obtained by the encoder. Therefore, not only global semantic information, but also

r
local semantic information is required. The sliding window in Swin-Transformer
computes self-attention in locally non-overlapping windows and allows cross-
window connections. Through windows multi-head self-attention (W-MSA) and

er
shifted windows multi-head self-attention (SW-MSA), Swin-Transformer can dy-
namically focus on adjacent elements and enhance the ability of local information
extraction, as well as the global attention ability.
pe
The entropy model based on discrete GMM achieves reasonable performance
in terms of rate-distortions, but during encoding and decoding, the probability dis-
tribution function (PDF) and cumulative distribution function (CDF) are dynam-
ically generated for each pixel, which results in redundancy and is time consum-
ing. In contrast, entropy models based on single gaussian model (SGM) construct
ot

fixed PDF and CDF tables for entropy encoding with lower computational cost.
To accelerate the serial processing steps in autoregressive context models, He [34]
proposes a parallelizable checkerboard context model that changes decoding order
to speed up decoding without affecting performance.
tn

Based on these considerations, this paper propoed an SGM based entropy

model named SwinTransEntropy for image compression with three features. First,
the model applies Swin-Transformer as the backbone of the codec module to
capture the correlation between spatially adjacent elements and global seman-
rin

tic information. Second, a fusion context autoregressive model is designed, and

checkerboard autoregressive is applied to speed up decoding. Third, the model
explores the decoupling of entropy parameters, that is, using different branches
to estimate entropy, which is helpful to improve the accuracy of entropy estima-
ep

tion. Finally, an asymmetric decoder is applied to further enhance the ability to

reconstruct clear textures and improve the reconstructed image quality. Extensive
experimental results show that our method outperforms state-of-the-art (SOTA)
image compression methods on fundamental metrics.
Pr

This preprint research paper has not been peer reviewed. Electronic copy available at: [Link]
 ed
3. Proposed SwinTransEntropy model

iew
3.1. Overall model structure
In this section, we present our lossy image compression scheme in detail.
The overall structure follows the general VAE based framework in Figure 1. The
schematic diagram of our method is illustrated in Figure 2.
Encoder
HyperEncoder
LinerEmbedding
PatchPartition

PatchMerging

PatchMerging
PatchMerging

ev
SwinͲTBlock

SwinͲTBlock

SwinͲTBlock
SwinͲTBlock

LeakyRelu
LeakyRelu
y

Conv2d
Conv2d

Conv2d

Conv2d
Conv2d
Conv2d z
FusionContextModel
Q
x x2 x2 x6 x2
Mask Mask Mask
yො Convolution
3*3
Convolution
5*5
Convolution
7*7

add Q

zො

r
AE AE

LeakyRelu

LeakyRelu

LeakyRelu
ConvT2d

ConvT2d

ConvT2d

ConvT2d
bitstream (ʅ漓ʍ) LeakyRelu concat bitstream

LeakyRelu

LeakyRelu
ConvT2d

ConvT2d

ConvT2d

ConvT2d
Decoder
AD

er HyperDecoder
AD
PatchExpanding

PatchExpanding

yො
PatchExpanding

zො
DeͲEmbedding

ResidualBlock
SwinͲTBlock

SwinͲTBlock

SwinͲTBlock
SwinͲTBlock
DeͲPartition

pe
LeakyRelu
LeakyRelu
Conv2d

Conv2d
Conv2d

Conv2d

Conv2d
Conv2d
xො x2 x2 x6 x2 x9

Figure 2: Schematic diagram of the proposed SwinTransEntropy model.

Denoting the original image as x. The image compression works as fol-

ot

lows. First, the latent representation feature y is obtained through an encoder,

i.e., y = Encoder(x), and then quantized into ŷ, i.e., ŷ = Q(y). Then ŷ is entropy
encoded and compressed into a bitstream. The hyperprior module and the context
tn

autoregressive module work together to obtain the entropy parameters for entropy
coding. Meanwhile, the side information z is extracted through a hyperencoder,
i.e., z = HyperE ncoder(y) and quantized into ẑ, which is entropy encoded and
compressed into a bitstream. Then ẑ is used to accurately estimate the entropy of
rin

the latent representation feature y using hyperdecoder and context autoregressive

module, returning entropy parameters µ and σ . Each pixel in ŷ is modelled as a
single Gaussian model with mean and variance as µ, σ :

pŷ|ẑ (ŷ | ẑ) = ∏ pŷ|ẑ (ŷi | ẑ) (1)

ep

ŷ is recoved from bitstream and entropy parameters. Finally, the image x̂ is recon-
structed via decoder, i.e., x̂ = Decoder(ŷ).
Pr

This preprint research paper has not been peer reviewed. Electronic copy available at: [Link]
 ed
3.2. Swin-Transformer based encoder

iew
Transformer was originally proposed in NLP fields and then extended to Vision-
Transformer for computer vision tasks. Vision-Transformer uses multi-head atten-
tion to improve the ability of focusing on multiple locations, and to model global
semantic information. Compared to Vision Transformer, Swin-Transformer is
more efficient and has a higher accuracy. Swin-Transformer uses a hierarchi-
cal construction method similar to convolutional neural networks. Through patch

ev
merging, extracted feature is down-sampled to increase the receptive field of the
next window attention on the original image, thus achieving a multi-scale feature
extraction capability.
The core module of Swin-Transformer is the Swin-Transformer block shown

r
in Figure 3. Windows multi-head self-attention (W-MSA) is to divide the fea-
ture map into multiple disjoint areas and Multi-Head Self-Attention is performed

er
in each window. As W-MSA limits the attention only to each window, Shifted
Windows Multi-Head Self-Attention (SW-MSA) is further applied that introduces
cross-connection between windows, making Swin-Transformer to obtain near-
global attention capability. For more information, see [19].
pe
MLP MLP
ot

LN LN
tn

W-MSA SW-MSA

LN LN
rin

Figure 3: Structural diagram of Swin-Transformer block.

By using Swin-Transformer block as the backbone of the codec module, the

ep

correlation and global semantic information between spatially adjacent elements

can be captured. The encoder used in our model is shown in Figure 4. The
patch partitioning module divides the input image into patches of size of K × K
which are then flattened over channel derection. For a RGB three-channel image
Pr

10

This preprint research paper has not been peer reviewed. Electronic copy available at: [Link]
 ed
with height H and width W , the shape of an image is changed from [H,W, 3] to
[H/K,W /K, K × K × 3]. Then, a linear transform is performed through the lin-

iew
ear embedding layer and the shape of the image is changed from [H/K,W /K,C].
Such a feature extraction process is repeated over different image sizes with four
stages. Stage1 is composed of a linear embedding layer and repeated stacked
Swin-Transformer block, and the remaining three stages are composed of patch
merging and repeated stacked Swin-Transformer block. Finally, the dimension of

ev
feature information is reduced by a convolution layer. To avoid too much spatial
redundant information in the patches, we set the patch size K to 2.

r
LinerEmbedding
PatchPartition

PatchMerging
PatchMerging

PatchMerging

SwinͲTBlock

SwinͲTBlock
SwinͲTBlock

SwinͲTBlock
InputImage

Conv2d
x2 x2
er x6 x2
pe
Figure 4: Structure diagram of encoder.

3.3. Entropy Model

The accuracy of the entropy estimation strongly affects the image compression
performance. In our method, we use a SGM-based autoregressive entropy model.
Building on the work of [7], we propose a fusion contextual autoregressive model
ot

to improve model prediction accuracy. In [7], the author experimentally proved

that 5*5 mask convolution is better than 3*3 and 7*7 mask convolution. Here, we
propose a fusion contextual autoregressive model shown in Figure 5, which fuses
tn

the tensors producted by 3*3, 5*5 and 7*7 mask convolution. We hope that the
additional context provided by 3*3 and 7*7 mask convolution can improve the
prediction accuracy. In addition, the entropy parameters, i.e., mean value µ and
variance σ , are decoupled in order to make the entropy estimation of potential
information more accurate and efficient.
rin

3.4. Asymmetric Decoder

Inspired by the work of Dosovitskiy [17] and He [35], we use an asymmetric
decoder as shown in Figure 6. First, the dimension is increased by a convolution
ep

layer. Then, the feature is up-sampled by stacking 9 residual blocks and 4 Swim-
Transformer blocks. By combining traditional CNN and Swin-Transformer, the
layer of the decoder module is increased so that the decoder has enough capacity
to learn the reconstruction ability.
Pr

11

This preprint research paper has not been peer reviewed. Electronic copy available at: [Link]
 Pr
ep
Conv2d Conv2d

ChannelNorm

x9
ResidualBlock

(ʅ,ʍ)
rin
Relu

x2
SwinͲTBlock
Conv2d
ConvT2d ConvT2d
PatchExpanding
ChannelNorm LeakyRelu LeakyRelu
tn
ConvT2d ConvT2d
+
LeakyRelu LeakyRelu

x6
SwinͲTBlock

ConvT2d ConvT2d
ot
PatchExpanding LeakyRelu LeakyRelu

12
ConvT2d ConvT2d
3*3
MaskConvolution

x2
SwinͲTBlock
pe
5*5

PatchExpanding
add

concat
MaskConvolution

er

x2

Figure 6: Structure diagram of asymmetric decoder.

SwinͲTBlock
7*7

r
MaskConvolution

DeͲEmbedding
Figure 5: Sturctural diagram of fusion context autoregressive model.

ev
DeͲPartition

Reconstruction
iew
ed

This preprint research paper has not been peer reviewed. Electronic copy available at: [Link]
 ed
3.5. Loss Function

iew
For model training, the loss function used is
L = λ d (x, x̂) + R(y) (2)
where, d (x, x0 ) is an image distortion measure, and Mean-Squared Loss (MSE) is
usually applied, R(y) is the bit rate of latent information y, expressed as
 

ev
R(y) = − log2p(ŷ|ẑ) (ŷ | ẑ) + log2 pẑ (ẑ) (3)

and λ is a parameter weighting image distortion and bit rate. Different λ will
results in different compression ratios.

r
4. Experimental results
4.1. Experimental setup
4.1.1. Datasets
er
The image compression task is an unsupervised task that does not require ad-
ditional label files. To validate the performance of our proposed methods, we
pe
conduct a series of experiments. We randomly sample 300K images from the
Open Images dataset, and randomly crop them to a size of 256×256 as training
set. Two public datasets, i.e., Kodak image set [36] and Tecnick SAMPLING im-
age set [37], are used for evaluation. It is generally believed that the performance
results on these qualified image test sets are representative and can be generalized
ot

to other untested natural images. The Kodak image set contains 24 lossless im-
ages with a resolution of 768 × 512. The Tecnick image set consists of 40 lossless
images with a resolution of 1200× 1200. In addition, to investigate the effective-
tn

ness of our proposal, ablation experiments are performed on two more datasets,
i.e, CLIC2020 validation set [38] and DIV2K validation set [39]. The CLIC2020
image set and The DIV2K image set use the validation set images, which consist
of 41 and 100 images, respectively.
rin

4.1.2. Compared methods

We compare the performance of our method with traditional image compres-
sion methods JPEG2000 [2], BPG [3], the latest H.266/VVC [40] and seven influ-
ential deep learning-based methods including hyperprior model (Ballé2018) [6],
ep

autoregressive model (Minnen) [7], autoregressive super prior model based on

GMM and attention mechanism (Cheng) [21], Context-adaptive entropy model
(Lee) [41] and Coarse-to-fine hyper-prior modelg (Hu) [26], continuous bitrate
adjustable model (Cui) [22] and Transformer-based model (Qian) [32],
Pr

13

This preprint research paper has not been peer reviewed. Electronic copy available at: [Link]
 ed
4.1.3. Trainging setting

iew
The parameter λ controls the bit rate, and we use a set of values {0.0018, 0.0035,
0.0067, 0.013, 0.025, 0.0485} to train our model, where each value corresponding
to a rate-distortion point. All models are trained for 1.2M steps using the Adam
optimizer [42] with a batch size of 16, The learning rate is set to 1 × 10−4 for 70K
iterations, and drops to 1 × 10−5 for the rest 50K iterations. We train our model
on a server with Intel Xeon Platinum 6164 CPU, NVIDIA 3090Ti GPU and 64GB

ev
RAM.

4.1.4. Evaluation index

Several metrics are used to evaluate model performance. Bits per pixel (BPP)

r
reflects the compression ratio of the image. For images with the same aspect ratio,
the smaller the BPP, the higher the compression ratio. Peak signal-to-noise ratio
(PSNR) and multiscale structural similarity index measure (MS-SSIM) are used

er
to evaluate image quality. PSNR is based on an element-wise difference compari-
son, which reflects pixel-level distortion. MS-SSIM measures image similarity in
terms of brightness, contrast, and structure, and is the most widely used percep-
pe
tual metric for evaluating (and training) deep learning-based image compression
models. The unit of PSNR and MS-SSIM is dB, and a higher value indicates bet-
ter performance. LPIPS [43], measures the distance in the feature space of a deep
neural network originally trained for image classification, is adapted for predict-
ing the similarity of distorted patches, which is validated to predict human scores
ot

for these distortions.

4.2. Comparison with SOTA methods

4.2.1. Rate-Distortion performance
tn

The RD curves using PSNR and MS-SSIM as distortion metrics are given in
Figure 7. It is noting that some methods are not considered in the comparison be-
cause the model parameters on a particular data set are not available. Figure 7(a)
gives the RD curves of each competing method on the Kodak Image Set under
rin

PSNR quality assessment. The average performance on the dataset shows that our
method (MSE-optimized) outperforms all the competing deep image compres-
sion methods optimized with the same metric. Figure 7(b) gives the RD curves of
each comparison method on Kodak under the MS-SSIM quality assessment. The
ep

results show that our MSE-optimized method outperforms other MSE-optimized

deep image compression methods and conventional codecs. Figures 7(c)-7(d) fur-
ther demonstrate the superior coding performance of our method on the Tecnick
Pr

14

This preprint research paper has not been peer reviewed. Electronic copy available at: [Link]
 ed
image set with PSNR and MS-SSIM metrics, and in general the results are simi-

iew
lar to the Kodak Image Collection. Therefore, we can conclude that our method
achieves the best performance among all the methods.

36 20

34 18

ev
32 16

MS-SSIM
PSNR

Proposed[MSE]
30 JPEG2000 14
BPG Proposed[MSE]
Balle[MSE](ICLR2018) JPEG2000
Minnen[MSE](NIPS2018) BPG
28 12
Hu[MSE](AAAI2020) Balle[MSE](ICLR2018)

r
Lee[MSE](ICLR2019) Minnen[MSE](NIPS2018)
Cheng[MSE](CVPR2020) Hu[MSE](AAAI2020)
26 Cui[MSE](CVPR2021) 10 Lee[MSE](ICLR2019)
Qian[MSE](ICLR2022) Cheng[MSE](CVPR2020)

0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8

40

39
Proposed[MSE]
JPEG2000
BPG
Bits Per Pixel

(a) er 22

21
Proposed[MSE]
JPEG2000
BPG
Bits Per Pixel

(b)
pe
Minnen[MSE](NIPS2018) Minnen[MSE](NIPS2018)
Hu[MSE](AAAI2020) Hu[MSE](AAAI2020)
Lee[MSE](ICLR2019) Lee[MSE](ICLR2019)
38 Cheng[MSE](CVPR2020) 20 Cheng[MSE](CVPR2020)
VVC-intra(VTM11.0) VVC-intra(VTM11.0)

37 19
MS-SSIM
PSNR

36 18

35 17

34 16
ot

33 15

32 14

0.05 0.10 0.15 0.20 0.25 0.30 0.35 0.05 0.10 0.15 0.20 0.25 0.30 0.35
tn

Bits Per Pixels Bits Per Pixels

(c) (d)

Figure 7: RD Performance evaluations on the Kodak and Tecnick dataset. (a) Kodak Image Set,
evaluated with MSE. (b) Kodak Image Set, evaluated with MS-SSIM. (c) Tecnick Image Set,
rin

evaluated with MSE. (d) Tecnick Image Set, evaluated with MS-SSIM.

4.2.2. Perceptual quality evaluation

To check the visual quality of reconstructed images, a popular perceptual qual-
ep

ity assessment method LPIPS is used. Lower LPIPS values indicate better visual
quality. In this experiments, we compare the visual quality of images recon-
structed by our method with Cheng, VVC intra profile, BPG and JPEG2000 at
Pr

15

This preprint research paper has not been peer reviewed. Electronic copy available at: [Link]
 ed
different rates. For a fair comparison with traditional image codecs, all evaluated

iew
codecs are optimized under MSE distortion. As shown in Figure 8(a), on the Ko-
dak image set, our proposed method achieves better perceptual quality than cheng,
JPEG2000, BPG, and VVC intra profiles. Figure 8(b) shows the LPIPS results on
the Tecnick Image Set. Our proposed method still achieves better visual quality
than Cheng, JPEG2000, and BPG, and comparable visual quality to VVC internal
profiles.

ev
0.45
Proposed[MSE] Proposed[MSE]
Cheng[MSE](CVPR2020) Cheng[MSE](CVPR2020)
0.40 VVC-intra(VTM11.0) 0.25 VVC-intra(VTM11.0)
BPG BPG
JPEG2000 JPEG2000
0.35

0.20

r
0.30
LPIPS

0.25 LPIPS
0.15
0.20

0.15

0.10
er 0.10
pe
0.05 0.05
0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.05 0.10 0.15 0.20 0.25 0.30 0.35
Bits Per Pixel Bits Per Pixel

(a) (b)

Figure 8: LPIPS Evaluation for MSE-optimized methods. (a) Kodak Image Set. (b) Tecnick Image
Set.
ot

4.2.3. Visual Effect

To make the effectiveness of our model clearer, we provide some visualiza-
tn

tion results. For fair comparison with conventional image codecs, all evaluated
codecs are optimized under MSE distortion. Since deep learning based image
compression methods cannot strictly limit the BPP of images, we can only com-
pare different compression methods when the BPP values are similar. Fig.9 gives
rin

the reconstructed images kodim19 from Kodak dataset generated by our method,
classical image compression codecs BPG and JPEG2000, and learning-based im-
age compression methods Cheng and Minnen. For the compared methods, we can
obsereve blurring artifacts. In contrast, our method can well recover image details
ep

and remove blur artifacts, verifying the effectiveness of our method.

4.3. Ablation Experiment
Pr

16

This preprint research paper has not been peer reviewed. Electronic copy available at: [Link]
 ed
iew
r ev
(a) Ground truth. (b) Proposed, bpp=0.119. (c) Cheng, BPP=0.129.

er
pe
ot

(d) Minnen, BPP=0.106. (e) BPG, BPP=0.235. (f) JPEG2000, BPP=0.215.

Figure 9: Image comparison of different compression methods with similar compression ratios
tn

(the picture is ”[Link]” from Kodak dataset).

4.3.1. Influence of patch size

To inverstigate the impact of patch size K on model performance, we train
rin

another model with K setting to 4 and embeding dimention setting to 96, with
network structure remaining unchanged. The network is trained under the same
training strategy and tested on Kodak dataset. We choose the models trained
with hyperparameter λ set to (0.0018, 0.0035, 0.0067). We plot the rate-distortion
ep

curves in Figure 10. According to the curves, patch size K settting to 2 is perfer-
able in terms of PSNR and MS-SSIM. The reason is a larger path size will bring
too much spatial redundant information for each window.
Pr

17

This preprint research paper has not been peer reviewed. Electronic copy available at: [Link]
 ed
15.0

iew
31.0
14.5
30.5
14.0
30.0
13.5

MS-SSIM
29.5
13.0
PSNR

29.0
12.5

28.5 12.0

ev
28.0 11.5
K=2[MSE] K=2[MSE]
27.5 K=4[MSE] 11.0 K=4[MSE]
0.08 0.10 0.12 0.14 0.16 0.18 0.20 0.22 0.24 0.26 0.08 0.10 0.12 0.14 0.16 0.18 0.20 0.22 0.24 0.26
Bits Per Pixel Bits Per Pixel

r
(a) (b)
Figure 10: The influence of patch size K in the proposed method.

4.3.2. Effectivenss of parameter decoupling

In our method, we decouple the mean and variance parameters of entropy
er
pe
model by using different branches to estimate their values. To verify the effective-
ness of such trick, we compare the performance with the version without using
decoupling. We use the model trained with hyperparameter λ set to 0.0035. The
two cases are quantitatively compared using PSNR and MS-SSIM on Kodak, Tec-
nick, CLIC2020 and DIV2K datasets, and the results are listed in Table 1. The
ot

results show that decoupling trick is benificial to further improve PSNR and MS-
SSIM while not sacrificing compression rate.
4.3.3. Effectivenss of fusion contextual autoregressive
tn

In [7], the author proved experimentally that 5*5 mask convolution is better
than 3*3 and 7*7 convolution. In our method we propose a fusion contextual
autoregressive model to improve model prediction accuracy, which fuses the ten-
sors output by 3*3, 5*5 and 7*7 mask convolution. To verify the effectiveness,
rin

we compare the performance with the version using single mask convolution, i.e.,
autoregressive model with 5*5 mask convolution. We use the model trained with
hyperparameter λ set to 0.0035. The two cases are quantitatively compared using
PSNR and MS-SSIM on Kodak, Tecnick, CLIC2020 and DIV2K datasets, and the
ep

results are listed in Table 2. The results show that our method can further improve
PSNR and MS-SSIM values to some extent, verifying our strategy is valid.
Pr

18

This preprint research paper has not been peer reviewed. Electronic copy available at: [Link]
 ed
iew
Table 1: Effectivenss of entropy parameters decoupling strategy.
With entropy Without entropy
Datasets Metric
parameters decoupling parameters decoupling
bpp 0.154 0.158
Kodak PSNR↑ 29.430 29.365

ev
MS-SSIM↑ 12.850 12.817
bpp 0.081 0.083
Tecnick PSNR↑ 33.809 33.732
MS-SSIM↑ 15.215 15.169

r
bpp 0.109 0.111
CLIC2020 PSNR↑ 31.390 31.323
MS-SSIM↑ 14.080 14.038

DIV2K
bpp
PSNR↑
MS-SSIM↑
er
0.150
30.260
14.330
0.154
30.189
14.289
pe

Table 2: Effectivenss of fusion contextual autoregressive.

ot

Fusion context Autoregressive model

Datasets Metric
autoregressive model with 5*5 mask conv
bpp 0.154 0.156
tn

Kodak PSNR↑ 29.430 29.360

MS-SSIM↑ 12.850 12.787
bpp 0.081 0.081
Tecnick PSNR↑ 33.809 33.662
MS-SSIM↑ 15.215 15.175
rin

bpp 0.109 0.109

CLIC2020 PSNR↑ 31.390 31.281
MS-SSIM↑ 14.080 14.014
bpp 0.150 0.151
ep

DIV2K PSNR↑ 30.260 30.151

MS-SSIM↑ 14.330 14.281
Pr

19

This preprint research paper has not been peer reviewed. Electronic copy available at: [Link]
 ed
4.3.4. Effectivenss of asymmetric decoder

iew
In our method, we combine the traditional CNN and Swin-Transformer to con-
truct an asymmetric decoder so that the decoder has enough capacity to learn the
reconstruction ability. To verify the effectiveness, we compare the performance
with the version using symmetric decoder. We also use the model trained with
hyperparameter λ set to 0.0035. The two cases are quantitatively compared using
PSNR and MS-SSIM on Kodak, Tecnick, CLIC2020 and DIV2K datasets, and

ev
the results are listed in Table3. According to the results, the use of asymmetric
decoder can also improve PSNR and MS-SSIM values, verifying its effectiveness.

Table 3: Effectivenss of asymmetric decoder.

r
Datasets Metric Asymmetric decoder Symmetric decoder
bpp 0.154 0.161
Kodak

Tecnick
PSNR↑
MS-SSIM↑
bpp
PSNR↑
er
29.430
12.850
0.081
33.809
29.330
12.840
0.084
33.554
pe
MS-SSIM↑ 15.215 15.148
bpp 0.109 0.112
CLIC2020 PSNR↑ 31.390 31.271
MS-SSIM↑ 14.080 14.056
bpp 0.150 0.155
ot

DIV2K PSNR↑ 30.260 30.148

MS-SSIM↑ 14.330 14.317
tn

5. Conclusion
In this paper, we propose a Swin-Transformer based entropy model for image
compression. The model utilizes a Swin-Transformer based encoder that can bet-
rin

ter capture correlations and global semantic information between spatially adja-
cent elements, decouples the mean and variance parameters entropy model and use
a fusion context autoregressive to obtain accurate entropy estimation and efficient
entropy coding, and uses an asymmetric decoder with more layers to increase the
ep

capacity of restoring texture information and reconstructing images. Experimen-

tal results show that the proposed method outperforms seven deep learning based
SOTA methods, and achieves high perceptual fidelity on reconstructing images.
Pr

20

This preprint research paper has not been peer reviewed. Electronic copy available at: [Link]
 ed
Next, we will further explore more efficient and lightweight feature extraction

iew
module for image compression tasks.

Acknowledgements
This work was supported by the National Natural Science Foundation for
Young Scientists of China (61603319, 61601385).

ev
References
[1] Gregory K Wallace. The jpeg still picture compression standard. Communi-
cations of the ACM, 34(4):30–44, 1991.

r
[2] David S Taubman, Michael W Marcellin, and Majid Rabbani. Jpeg2000:
Image compression fundamentals, standards and practice. Journal of Elec-
tronic Imaging, 11(2):286–287, 2002.
er
[3] Bellard Fabrice. BPG image format. [Link] 2014.
pe
[4] Nasir Ahmed, T Natarajan, and Kamisetty R Rao. Discrete cosine trans-
form. IEEE transactions on Computers, 100(1):90–93, 1974.
[5] Diederik P Kingma and Max Welling. Auto-encoding variational bayes.
arXiv preprint arXiv:1312.6114, 2013.
ot

[6] Johannes Ballé, David Minnen, Saurabh Singh, Sung Jin Hwang, and Nick
Johnston. Variational image compression with a scale hyperprior. arXiv
preprint arXiv:1802.01436, 2018.
tn

[7] David Minnen, Johannes Ballé, and George D Toderici. Joint autoregressive
and hierarchical priors for learned image compression. Advances in neural
information processing systems, 31, 2018.
rin

[8] Detlev Marpe, Heiko Schwarz, and Thomas Wiegand. Context-based adap-
tive binary arithmetic coding in the h. 264/avc video compression standard.
IEEE Transactions on circuits and systems for video technology, 13(7):620–
636, 2003.
ep

[9] George Toderici, Sean M O’Malley, Sung Jin Hwang, Damien Vincent,
David Minnen, Shumeet Baluja, Michele Covell, and Rahul Sukthankar.
Variable rate image compression with recurrent neural networks. arXiv
preprint arXiv:1511.06085, 2015.
Pr

21

This preprint research paper has not been peer reviewed. Electronic copy available at: [Link]
 ed
[10] Johannes Ballé, Valero Laparra, and Eero P Simoncelli. End-to-end opti-

iew
mized image compression. arXiv preprint arXiv:1611.01704, 2016.
[11] Lucas Theis, Wenzhe Shi, Andrew Cunningham, and Ferenc Huszár.
Lossy image compression with compressive autoencoders. arXiv preprint
arXiv:1703.00395, 2017.
[12] Volodymyr Mnih, Nicolas Heess, Alex Graves, et al. Recurrent models of

ev
visual attention. Advances in neural information processing systems, 27,
2014.
[13] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In Pro-

r
ceedings of the IEEE conference on computer vision and pattern recognition,
pages 7132–7141, 2018.

er
[14] Zilong Zhong, Zhong Qiu Lin, Rene Bidart, Xiaodan Hu, Ibrahim Ben Daya,
Zhifeng Li, Wei-Shi Zheng, Jonathan Li, and Alexander Wong. Squeeze-
and-attention networks for semantic segmentation. In Proceedings of the
IEEE/CVF conference on computer vision and pattern recognition, pages
pe
13065–13074, 2020.
[15] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones,
Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you
need. Advances in neural information processing systems, 30, 2017.
ot

[16] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier,

Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with
transformers. In Computer Vision–ECCV 2020: 16th European Conference,
tn

Glasgow, UK, August 23–28, 2020, Proceedings, Part I 16, pages 213–229.
Springer, 2020.
[17] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn,
rin

Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer,

Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Trans-
formers for image recognition at scale. arXiv preprint arXiv:2010.11929,
2020.
ep

[18] Robin Strudel, Ricardo Garcia, Ivan Laptev, and Cordelia Schmid. Seg-
menter: Transformer for semantic segmentation. In Proceedings of the
IEEE/CVF international conference on computer vision, pages 7262–7272,
2021.
Pr

22

This preprint research paper has not been peer reviewed. Electronic copy available at: [Link]
 ed
[19] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen

iew
Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer
using shifted windows. In Proceedings of the IEEE/CVF international con-
ference on computer vision, pages 10012–10022, 2021.

[20] Johannes Ballé, Valero Laparra, and Eero P Simoncelli. End-to-end opti-
mization of nonlinear transform codes for perceptual quality. In 2016 Picture
Coding Symposium (PCS), pages 1–5. IEEE, 2016.

ev
[21] Zhengxue Cheng, Heming Sun, Masaru Takeuchi, and Jiro Katto. Learned
image compression with discretized gaussian mixture likelihoods and atten-
tion modules. In Proceedings of the IEEE/CVF Conference on Computer

r
Vision and Pattern Recognition, pages 7939–7948, 2020.

[22] Ze Cui, Jing Wang, Shangyin Gao, Tiansheng Guo, Yihui Feng, and Bo Bai.

er
Asymmetric gained deep image compression with continuous rate adapta-
tion. In Proceedings of the IEEE/CVF Conference on Computer Vision and
Pattern Recognition, pages 10532–10541, 2021.
pe
[23] Jooyoung Lee, Seunghyun Cho, and Munchurl Kim. An end-to-end joint
learning scheme of image compression and quality enhancement with im-
proved entropy minimization. arXiv preprint arXiv:1912.12817, 2019.

[24] David Minnen and Saurabh Singh. Channel-wise autoregressive entropy

ot

models for learned image compression. In 2020 IEEE International Confer-

ence on Image Processing (ICIP), pages 3339–3343. IEEE, 2020.

[25] Johannes Ballé, Valero Laparra, and Eero P Simoncelli. Density modeling
tn

of images using a generalized normalization transformation. arXiv preprint

arXiv:1511.06281, 2015.

[26] Yueyu Hu, Wenhan Yang, and Jiaying Liu. Coarse-to-fine hyper-prior mod-
rin

eling for learned image compression. In Proceedings of the AAAI Conference

on Artificial Intelligence, pages 11013–11020, 2020.

[27] Haojie Liu, Tong Chen, Peiyao Guo, Qiu Shen, Xun Cao, Yao Wang, and
Zhan Ma. Non-local attention optimized deep image compression. arXiv
ep

preprint arXiv:1904.09757, 2019.

[28] Sharon Ayzik and Shai Avidan. Deep image compression using decoder side
information. In Computer Vision–ECCV 2020: 16th European Conference,
Pr

23

This preprint research paper has not been peer reviewed. Electronic copy available at: [Link]
 ed
Glasgow, UK, August 23–28, 2020, Proceedings, Part XVII 16, pages 699–

iew
714. Springer, 2020.
[29] Fabian Mentzer, George D Toderici, Michael Tschannen, and Eirikur
Agustsson. High-fidelity generative image compression. Advances in Neural
Information Processing Systems, 33:11913–11924, 2020.
[30] Shangyin Gao, Yibo Shi, Tiansheng Guo, Zhongying Qiu, Yunying Ge,

ev
Ze Cui, Yihui Feng, Jing Wang, and Bo Bai. Perceptual learned image com-
pression with continuous rate adaptation. 2023.
[31] Ge Gao, Pei You, Rong Pan, Shunyuan Han, Yuanyuan Zhang, Yuchao Dai,
and Hojae Lee. Neural image compression via attentional multi-scale back

r
projection and frequency decomposition. In Proceedings of the IEEE/CVF
International Conference on Computer Vision, pages 14677–14686, 2021.

er
[32] Yichen Qian, Ming Lin, Xiuyu Sun, Zhiyu Tan, and Rong Jin. Entroformer:
A transformer-based entropy model for learned image compression. arXiv
preprint arXiv:2202.05492, 2022.
pe
[33] Zongyu Guo, Yaojun Wu, Runsen Feng, Zhizheng Zhang, and Zhibo Chen.
3-D context entropy model for improved practical image compression. In
Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern
Recognition Workshops, pages 116–117, 2020.
ot

[34] Dailan He, Yaoyan Zheng, Baocheng Sun, Yan Wang, and Hongwei Qin.
Checkerboard context model for efficient learned image compression. In
Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern
tn

Recognition, pages 14771–14780, 2021.

[35] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross
Girshick. Masked autoencoders are scalable vision learners. In Proceedings
of the IEEE/CVF Conference on Computer Vision and Pattern Recognition,
rin

pages 16000–16009, 2022.

[36] Eastman Kodak. Kodak lossless true color image suite(photocd pcd0992).
[Link] 1993.
ep

[37] Nicola Asuni and Andrea Giachetti. TESTIMAGES: a Large-scale Archive

for Testing Visual Devices and Basic Image Processing Algorithms. In An-
drea Giachetti, editor, Smart Tools and Apps for Graphics - Eurographics
Italian Chapter Conference. The Eurographics Association, 2014.
Pr

24

This preprint research paper has not been peer reviewed. Electronic copy available at: [Link]
 ed
[38] Radu Timofte et. al. George Toderici, Wenzhe Shi. Workshop and challenge

iew
on learned image compression (clic2020), 2020.

[39] Eirikur Agustsson and Radu Timofte. Ntire 2017 challenge on single im-
age super-resolution: Dataset and study. In Proceedings of the IEEE confer-
ence on computer vision and pattern recognition workshops, pages 126–135,
2017.

ev
[40] Benjamin Bross, Jianle Chen, Jens-Rainer Ohm, Gary J. Sullivan, and Ye-
Kui Wang. Developments in international video coding standardization after
avc, with an overview of versatile video coding (vvc). Proceedings of the
IEEE, 109(9):1463–1493, 2021.

r
[41] Jooyoung Lee, Seunghyun Cho, and Seung-Kwon Beack. Context-adaptive
entropy model for end-to-end optimized image compression. arXiv preprint
arXiv:1809.10452, 2018.
er
[42] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic opti-
pe
mization. arXiv preprint arXiv:1412.6980, 2014.

[43] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver
Wang. The unreasonable effectiveness of deep features as a perceptual met-
ric. In Proceedings of the IEEE conference on computer vision and pattern
recognition, pages 586–595, 2018.
ot
tn
rin
ep
Pr

25

This preprint research paper has not been peer reviewed. Electronic copy available at: [Link]