Deep residual neural networks for inverse halftoning

TELKOMNIKA Telecommunication Computing Electronics and Control
Vol. 20, No. 6, December 2022, pp. 1326~1335
ISSN: 1693-6930, DOI: 10.12928/TELKOMNIKA.v20i6.24230  1326
Journal homepage: http://telkomnika.uad.ac.id
Deep residual neural networks for inverse halftoning
Heri Prasetyo1
, Muhamad Aditya Putra Anugrah1
, Alim Wicaksono Hari Prayuda2
, Chih-Hsien Hsia3
1
Department of Informatics, Faculty of Mathematics and Natural Sciences, Universitas Sebelas Maret (UNS), Surakarta, Indonesia
2
Department of Electrical Engineering, College of Electrical Engineering and Computer Science, National Taiwan University of Science
and Technology, Taipei City, Taiwan
3
Department of Computer Science and Information Engineering, College of Electrical Engineering and Computer Science,
National Ilan University, Yilan County, Taiwan
Article Info ABSTRACT
Article history:
Received Apr 09, 2021
Revised Jul 30, 2022
Accepted Aug 15, 2022
This paper presents a simple technique to perform inverse halftoning using
the deep learning framework. The proposed method inherits the usability and
superiority of deep residual learning to reconstruct the halftone image into
the continuous-tone representation. It involves a series of convolution
operations and activation function in forms of residual block elements.
We investigate the usage of pre-activation function and standard activation
function in each residual block. The experimental section validates the
proposed method ability to effectively reconstruct the halftone image.
This section also exhibits the proposed method superiority in the inverse
halftoning task compared to that of the handcrafted feature schemes and
former deep learning approaches. The proposed method achieves 30.37 dB
and 0.9481 on the average peak signal-to-noise ratio (PSNR) and structural
similarity index (SSIM) scores, respectively. It gives the improvements
around 1.67 dB and 0.0481 for those values compared to the most competing
scheme.
Keywords:
Deep learning
Error diffusion
Floyd-steinberg
Inverse halftoning
Residual networks
This is an open access article under the CC BY-SA license.
Corresponding Author:
Heri Prasetyo
Department of Informatics, Faculty of Mathematics and Natural Sciences
Universitas Sebelas Maret (UNS), 36 Ir. Sutami Road, Kentingan, Jebres, Surakarta City 57126, Indonesia
Email: heri.prasetyo@staff.uns.ac.id
1. INTRODUCTION
The digital halftoning technique is powerful tools in the rendering devices such as the digital
printing [1]. It converts an image with continuous-tone representations into only two-tone values. This
conversion always considers the quality of rendered image in two-tone representations. Each pixel in the
rendered image, or called as halftone image, only has two values, i.e. black or white pixels. The combinations
of all pixels in black and white can generate visual illusion such that one may recognize the halftone image as
the original image under human perception over a specific distance view. A good digital halftoning is able to
produce the halftone image with similar visual illusion compared to the original input image.
Some applications require the conversion of the halftone image consisting of two-tone values back
into its continuous-tone image. This conversion process is referred as inverse halftoning. Numerous works
have been presented in order to develop a new method or to gain an increased performance for effective
inverse halftoning technique. Commonly, the former inverse halftoning methods are based on the handcrafted
features [2]–[6]. However, some advanced progresses have been made in the inverse halftoning task under the
deep learning approaches [7]–[9]. As reported in literature, the deep learning-based methods deliver better
results compared to that of the handcrafted feature. The deep learning frameworks have also been successfully
reported to produce sophisticated results in some applications, such as handwriting recognition [10],
road recognition [11], image reconstruction [12], impulsive noise suppression [13], sound event detection [14],

TELKOMNIKA Telecommun Comput El Control 
Deep residual neural networks for inverse halftoning (Heri Prasetyo)
1327
and handwriting identification [15]. The presented method in this paper performs inverse halftoning task
under the residual learning framework. This residual learning method can be further extended to several
applications such as image forgery detection [16], image enhancement [17], automatic image retrieval [18],
and other computer vision tasks. This paper is composed with the following organizations: section 2 briefly
discusses the digital halftoning technique using the error diffusion approach. This section also presents
several techniques for inverse halftoning. The proposed method for inverse halftoning using deep residual
learning is delivered in section 3. Section 4 discusses and summarizes some inverse halftoning experiments.
The conclusion remarks are then given at the of this paper.
2. IMAGE HALFTONING AND ITS INVERSE PROBLEM
The section briefly reviews the digital halftoning technique using the error diffusion approach [1].
This halftoning method utilizes a specific error kernel, namely floyd-steinberg kernel, to perform the image
thresholding approach. Subsequently, several methods for performing inverse halftoning are also presented in
this section. We firstly begin our discussion with the error diffusion using the floyd-steinberg kernel for two
dimensional grayscale image. This method can be easily extended for color image by treating each color
channel as an individual two dimensional grayscale image. Let 𝐹(𝑥, 𝑦) be an image pixel at position (𝑥, 𝑦)
for 𝑥 = 1, 2, … , 𝑊 and 𝑦 = 1, 2, … , 𝐻, where 𝑊 and 𝐻 denote the weight and height of an image,
respectively. This pixel is regarded as continues-tone since it lies in an interval [0,255]. The main goal of
digital halftoning is to convert this continues-tone pixel into two-tone presentations under the constraint that the
halftoned image gives almost similar visual appearance compared to the original continues-tone. For performing
the digital halftoning, one needs to compute the mean value over all image pixels using the (1).
𝜇 = ∑ ∑ 𝐹(𝑥, 𝑦)
𝐻
𝑦=1
𝑊
𝑥=1 (1)
Where 𝜇 is the average value of all pixels. Afterwards, the hard thresholding applies for each continues tone
pixel with:
𝐼(𝑥, 𝑦) = {
0, 𝐹(𝑥, 𝑦) < 𝜇
255, 𝐹(𝑥, 𝑦) ≥𝜇
(2)
Where 𝐼(𝑥, 𝑦) denotes the thresholded pixel at position (𝑥, 𝑦). Each thresholded pixel constitutes into a
single image, namely halftone image. The thresholding technique in (2) simply classifies a pixel which is
higher than mean value into the bright tone (white pixel), and vice versa.
In fact, the hard thresholding as defined in (2) produces an error, i.e. the difference between the
pixel 𝐹(𝑥, 𝑦) and 𝐼(𝑥, 𝑦). The error caused by the thresholding procedure can be compute as:
𝑒(𝑥, 𝑦) = 𝐹(𝑥, 𝑦) − 𝐼(𝑥, 𝑦) (3)
Where 𝑒(𝑥, 𝑦) represents the error value at pixel position (𝑥, 𝑦). The quality of halftone image 𝐼(𝑥, 𝑦) will be
improved and more acceptable if we are able to diffuse this error into its neighboring pixel. To conduct this
diffusion process, we need an auxiliary kernel, namely floy steinberg error kernel as shown in Figure 1.
The error diffusion process can be performed as:
𝐹(𝑥, 𝑦) ← 𝐹(𝑥, 𝑦) + 𝑒(𝑥, 𝑦) × 𝜀 (4)
Where 𝜀 and × are the value of error kernel and convolution operator, respectively. All thresholding and
error diffusion processes are repeated over all pixels to obtain the halftone image 𝐼. Figure 2 illustrates the
digital halftoning for color image. Herein, each color channel is independently processed with error diffusion
halftoning using floyd-steinberg kernel.
In most common situations, the human vision cannot fully accept the quality of halftone image in
dot-disperse and noisy-like patterns. It is caused by the truncated continuous-tone into only two tone
representations. Several efforts have been undertaken for improving the quality of halftone image or
reconstructing the halftone image back into its original continuous tone. Typically, a technique aiming to
convert the halftone image into continuous-tone is referred as inverse halftoning method. Figure 3 shows the
process of inverse halftoning. The methods [2]–[8] perform the inverse halftoning effectively by producing
high quality of reconstructed continuous tone. A simple approach for inverse halftoning is presented in [2]
which employs the maximum a posteriori. Whereas, the method in [4] executes the anisotropic diffusion to
reconstruct the halftone image. The technique in [3] improves the quality of halftone image using the
deconvolution operation and regularized wienner inverse, while the methods in [5], [6] exploit the superiority
of sparse representation technique to replace the halftone image patch with the continue-tone patch. All these

 ISSN: 1693-6930
TELKOMNIKA Telecommun Comput El Control, Vol. 20, No. 6, December 2022: 1326-1335
1328
techniques can be regarded as handcrafted inverse halftoning method. In recent years, some progresses and
developments have been achieved in the inverse halftoning task incorporating the recent advances of deep
learning approach. The method in [7] performs the inverse halftoning using deep learning approach under
the encoder-decoder computation, whereas former scheme in [8] takes the structure-aware deep learning
for inverse halftoning. All these mentioned methods yield promising results in the reconstructed halftoning
image. The halftoning method has been proven to yield an effective result on image retrieval [19] and
image compression [20].
Figure 1. Floy-steinberg error kernel Figure 2. Illustration of digital halftoning
Figure 3. Illustration of inverse halftoning
3. PROPOSED INVERSE HALFTONING
3.1. Proposed architecture
The proposed method exploits the residual learning to perform the inverse halftoning task. Let 𝐼 be the
halftone image. The proposed method aims to learn an efficient mapping that transforms a halftone image 𝐼 into
its continuous approximation, i.e. the reconstructed halftone image 𝐼
̂. This process is denoted as 𝐼
̂ = ℑ{𝐼}, while
ℑ{∙} is the proposed end-to-end mapping. The proposed method should consider the requirement that the quality
of 𝐼
̂ should be as similar as possible to 𝐼, i.e. 𝐼
̂ = 𝐼 or 𝐼
̂ ≈ 𝐼.
Suppose that 𝐼 be the halftone image in color space, while its size is denoted as 𝐻 × 𝑊 × 𝐶. Herein,
the image height, width, and the number of color channels are represented as 𝐻, 𝑊, and 𝐶, respectively.
The value of 𝐶 = 3 indicates the color image, whereas 𝐶 = 1 is grayscale image. Figure 4 depicts the
proposed method for inverse halftoning task. It receives an input image of size 𝐻 × 𝑊 × 𝐶, afterwards,
it produces the output image over identical size with the original image. Firstly, the proposed method
performs the convolution operation on 𝐼 using the (5).
𝐼𝑎 = 𝜂(𝐼 × 𝑊
𝑎 + 𝑏𝑎) (5)
Where 𝑊
𝑎 and 𝑏𝑎 are the weights and biases, respectively, in this convolution process. This convolution
process consists of 𝑛𝑎 filters, where each filter is of size 𝑓𝑎 × 𝑓𝑎 × 𝐶. The symbol 𝜂(∙) is the activation
function. This convolution process produces 𝑛𝑎 feature maps denoted as 𝐼𝑎.
Figure 4. Schematic diagram of the proposed resnet architecture for inverse halftoning
Halftone Image
Conv2D
Activation
ResBlock[1]
ResBlock[2]
ResBlock[D]
Identity Mapping
Conv2D
Activation
Conv2D
Continuous tone Image

1329
The feature maps produced at the first convolution layer are subsequently fed into a series of
residual blocks. The process of each residual block is formally defined:
𝐼𝑑 = 𝑅𝑑{𝐼𝑑−1} (6)
Where 𝑅𝑑{∙} is the operator in the 𝑑-th residual block, for 𝑑 = 1, 2, … , 𝐷. The symbol 𝐷 indicates the number
of residual blocks. While 𝐼𝑑 denotes the feature maps produced at the 𝑑-th residual block. For 𝑑 = 1, we simply
set 𝐼0 = 𝐼𝑎. The proposed method produces the feature maps 𝐼𝐷 at the end of residual blocks, i.e. 𝑑 = 𝐷.
The dimensionality of feature maps 𝐼𝐷 should be identically maintained as in the feature maps 𝐼𝑎, i.e. 𝑛𝑎. Since,
we perform identity mapping 𝐼𝑎 and element-wise addition at the end of residual blocks [21]. The process of
element-wise addition is denoted as:
𝐼𝑏 = 𝐼𝑎⨁𝐼𝐷 (7)
Where 𝐼𝑏 and ⊕ denote the feature maps after element-wise addition, and operator of this addition,
respectively. The dimensionality of 𝐼𝑏 is also 𝑛𝑎.
After element-wise addition process, the two convolution series are applied to the feature maps 𝐼𝑏.
The first convolution process of 𝐼𝑏 is indicated with the:
𝐼𝑐 = 𝜂(𝐼𝑏 × 𝑊
𝑐 + 𝑏𝑐) (8)
Where 𝑊
𝑐 and 𝑏𝑐 are the weights and biases, respectively, for this convolution. This process involves 𝑛𝑐
filters, each of size 𝑓𝑐 × 𝑓𝑐 × 𝑛𝑎. This process yields the feature maps 𝐼𝑐 of size 𝑛𝑐. This process utilizes the
activation function. Subsequently, the second process performs the convolution of 𝐼𝑐 as:
𝐼𝑒 = 𝐼𝑐 × 𝑊
𝑒 + 𝑏𝑒 (9)
Where 𝑊
𝑒 and 𝑏𝑒 are the weights and biases, respectively, involved in the convolution operation. This stage
does not involve the activation function. In addition, this process needs 𝑛𝑒 filters, each of size 𝑓𝑒 × 𝑓𝑒 × 𝑛𝑐 to
yield feature maps 𝐼𝑑 of size 𝑛𝑒. By setting 𝑛𝑒 = 𝐶, we obtain the final feature maps 𝐼𝑒 with an identical size to
that of the original input image. We denote 𝐼𝑒 as the reconstructed halftone image, i.e. 𝐼
̂ = 𝐼𝑒. We set 𝑛𝑒 = 𝐶 = 3
for color image, while 𝑛𝑒 = 𝐶 = 1 for grayscale image.
3.2. Elements of normal residual block
This subsection explains the elements of each residual block used in the proposed method.
As discussed before, each residual block performs feature mapping as 𝐼𝑑 = 𝑅(𝐼𝑑−1) for 𝑑 = 1, 2, … , 𝐷. In the
normal residual block setting, we simply employ two convolution series. The illustration of this residual
block is given in Figure 5. The first convolution is denoted as:
𝑀𝑎 = 𝜂{𝑁𝑜𝑟𝑚{𝐼𝑑−1 × 𝑊1 + 𝑏1}} (10)
Where 𝑊1 and 𝑏1 are the weights and biases of this convolution process, respectively. The symbol 𝑁𝑜𝑟𝑚{∙}
denotes the operator of image batch normalization [22]. This stage produces the feature maps 𝑀𝑎. The second
convolution processes the feature maps 𝑀𝑎 using the:
𝑀𝑏 = 𝜂{𝑁𝑜𝑟𝑚{𝑀𝑎 × 𝑊2 + 𝑏2}} (11)
Where 𝑊2 and 𝑏2 are the weights and biases in this convolution process. This process produces the feature
maps 𝑀𝑏. The element-wise addition is performed at the end of each residual block as:
𝑀𝑐 = 𝑀𝑎 ⊕ 𝑀𝑏 (12)
Where 𝑀𝑐 is the produced feature maps implying 𝐼𝑑 = 𝑀𝑐. These two convolution layers are applied to all
residual blocks for 𝑑 = 1, 2, … , 𝐷.
3.3. Elements of residual block with pre-activation function
The residual block can be alternatively designed with the pre-activation function. In this scenario,
the activation function is firstly executed before the convolution process. The process of each residual block
is denoted as 𝐼𝑑 = 𝑅(𝐼𝑑−1) for 𝑑 = 1,2, … , 𝐷. The residual block with pre-activation function also involves
two convolution series. Figure 6 shows the element of residual block with pre-activation function. The first
convolution process is performed as:

 ISSN: 1693-6930
1330
𝑀𝑎 = 𝜂{𝑁𝑜𝑟𝑚{𝐼𝑑−1}} × 𝑊1 + 𝑏1 (13)
Whereas, the second convolution process on this residual block is conducted as:
𝑀𝑏 = 𝜂{𝑁𝑜𝑟𝑚{𝑀𝑎}} × 𝑊2 + 𝑏2 (14)
It can be seen from these two formulations, the activation function is applied before convolution operation.
The element-wise operation is further computed as:
𝑀𝑐 = 𝑀𝑎 ⊕ 𝑀𝑏 (15)
The feature maps produced at this residual block are denoted as 𝑀𝑐 or 𝐼𝑑 = 𝑀𝑐. This process is performed for
all residual blocks 𝑑 = 1,2, … , 𝐷.
Figure 5. Elements of normal residual block Figure 6. Elements of residual block with pre-activation
function
3.4. Loss function in learning process
The proposed method performs end-to-end learning to reconstruct the halftone image. This process
involves the training activity under a specific loss function. The mean-squared error (MSE) is chosen as our
loss function since of its simplicity. The MSE definition is given as:
𝑙(𝐼
̂, 𝐹) = ‖𝐼
̂ − 𝐹‖𝐹
2
(16)
Where 𝑙(𝐼
̂, 𝐹) denotes the loss function between the reconstructed image 𝐼
̂ and original image 𝐹. The symbol
‖∙‖𝐹 is the operator of frobenius norm. To reduce the computational time, we conduct the training process on
the image patch basis rather than the full size image. Let {𝐼𝑡, 𝐹𝑡}𝑡=1
𝑁
be a paired halftone image patch and its
original version, while 𝑁 denotes the number of image patches involved in the training process.
Let Θ = {𝑊
𝑎, 𝑏𝑎, 𝑊𝑏, 𝑏𝑏, … , 𝑊
𝑒, 𝑏𝑒} is the trainable parameters for the proposed method. Subsequently,
the optimization process can be easily performed under the image patch basis as:
𝑎𝑟𝑔 𝑚𝑖𝑛
𝛩
1
𝑁
∑ 𝑙((𝐼𝑡, 𝛩), 𝐹𝑡)
𝑁
𝑡=1 = 𝑎𝑟𝑔 𝑚𝑖𝑛
𝛩
1
𝑁
∑ ‖(𝐼𝑡, 𝛩) − 𝐹𝑡‖𝐹
2
𝑁
𝑡=1 (17)
Where symbol (𝐼𝑡, Θ) denotes the inverse halftoning on image patch 𝐼𝑡 with the network parameters 𝛩. This
optimization requires a number of iterations to yield optimum results.
4. EXPERIMENTAL RESULTS
4.1. Parameters tuning
In this subsection, we perform parameters tuning for our proposed resnet method. The parameters
tuning is performed in the training process. Herein, we need two image sets in color version. The first image
set is referred as training set obtained from the DIV2K image dataset [23]. Each image is divided into
non-overlapping image patches, each of size 128×128. Figure 7 displays some training image samples.
Whereas, the second image set is the validation set composed from DIV2K image dataset [23] in the
downsampled version with bicubic interpolation by scale × 4. Figure 8 exhibits some image samples for the
validation set. We need a paired images to train the proposed network, i.e. clean (original) image and
halftone image version. In this stage, each halftone image is fed into the proposed network, then the error
between the reconstructed image produced by the proposed network is computed against the clean image.
All experiments are conducted under the computational environments: AMD Ryzen threadripper 1950X CPU
and Nvidia GTX 1080 Ti GPU.
Conv2D
Activation
Elementwise
Addition
Normalization
Conv2D
Activation
Normalization
Conv2D
Activation
Elementwise
Addition
Normalization
Conv2D
Activation
Normalization

1331
Figure 7. Some image samples as training set
Figure 8. A set of validation images
We firstly overlook the effect of different numbers of features for the proposed network. Herein,
we simply use the MSE metric as a loss function. We employ an image batch with size 32. The Adam
optimizer [24] performs optimization of network parameters with initial learning rate 0.001. This learning
value is further divided with 2 on every 5 epochs. The number of epoches is set as 30, while each epoch
consists of 1261 iteration. All weights and biases are initialized using the similar strategy as used in [25].
In this experiment, we simply use four residual blocks without batch normalization. The parametric rectified
linear unit (PreLU) is employed as an activation function.
Figure 9(a) shows the average loss function in terms of peak signal-to-noise ratio (PSNR) score over
validation image set under various number of feature maps, i.e. 𝑁 = 16, 𝑁 = 32, and 𝑁 = 48. As depicted
from this figure, the 48 feature maps yield the best performance in the training process. Thus, we simply use
the feature maps as 𝑁 = 48 for the subsequent experiment.
Subsequently, we investigate the effect of various residual blocks for the proposed network. Herein,
we observe different number of residual blocks 𝐷 = {2, 4, … , 10} under the number of feature maps 𝑁 = 48.
Whereas, the other experimental settings remain unchanged. Figure 9(b) exhibits the performance of the
proposed network during the training stage over various number of residual blocks. This figure reveals that
the number of residual blocks 𝐷 = 10 gives the best performance for the proposed method indicating with
the highest average PSNR score in the validation set. Thereafter, we utilize 𝑁 = 48 and 𝐷 = 10 for the
subsequent training process.
The effects of activation function and image batch normalization method are examined for the
proposed network. In this experiment, we employ 𝑁 = 48 and 𝐷 = 10, while the other parameters setting is
maintained unchanged. Several activation functions such as PReLU, rectified linear unit (RELU), and others are
investigated. Figure 9(c) reports the average loss of validation set over various activation functions. This figure
indicates that the PReLU activation function offers the best performance compared to the other functions.
As a result, the PReLU activation function is more preferable compared to the others.
The effect of image batch normalization is further observed under 𝑁 = 48, 𝐷 = 10, and PReLU
activation function. Herein, we investigate the effect of batch normalization (BN) and instance normalization
(IN) with per-activation and normal activation setting. Figure 9(d) displays the effect of various batch
normalization for the proposed method. Based on this figure, we can conclude that the proposed method
without image batch normalization gives the best performance for validation set during the training process.
Thus, we apply the proposed method with experimental settings 𝑁 = 48, 𝐷 = 10, PReLU activation function
and without image batch normalization for the inverse halftoning task.
4.2. Visual investigation
This subsection inspects the performance under visual investigation. The quality of reconstructed
image is visually noticed and compared with the original image. Herein, we examine the performance under
several images as previously used in [6]. Figure 10 displays some color images for the testing set. These
testing images contain rich image detail and texture making it challenging in the inverse halftoning task.
For performing the inverse halftoning, we utilize the proposed method with the best parameters setting as

 ISSN: 1693-6930
1332
described in section 4.1. Figure 11 exhibits the visual comparisons on the reconstructed halftone images.
As shown in these figures, the proposed method produces the best visual quality of reconstructed image in
comparison to other scheme [6]. Since of the paper length limitation, we only give visual comparison against [6].
Therefore, it offers competitive advantage in the inverse halftoning task.
(a) (b)
(c) (d)
Figure 9. The average PSNR during the training process over various conditions: (a) feature maps,
(b) the number of residual blocks, (c) activation function, and (d) batch normalization
Figure 10. A set of testing images used for experiments
Original Halftone LLDO [6] Reconstructed
Figure 11. Visual investigation on bear image

1333
4.3. Performance comparisons
Additional comparisons are summarized and discussed in this subsection. We compare the
performances in terms of objective image quality assessments, i.e. the average PSNR and structural similarity
index (SSIM) scores over testing set as given in Figure 10. The inverse halftoning technique is applied for
each image in the testing set. Subsequently, the average PSNR and SSIM values are computed over all
testing images. We employ the proposed method with the best parameters tuning as discussed in section 4.1.
Table 1 tabulates the performance comparisons in terms of average PSNR, while Table 2 recapitulates these
comparisons in terms of average SSIM score. As tabulated in these two tables, the proposed method
outperforms the former existing schemes in the inverse halftoning task indicated with the highest values of
PSNR and SSIM. Herein, the proposed method and former schemes employ an identical experimental
environments, i.e. use identical image dataset, utilize the same floyd-steinberg halftoning technique, and
measure the performance under the similar image assessment metrics. The proposed method yields a good
performance in the inverse halftoning task, not only under visual investigation, but it also proves its
outstanding performance in the objective measurement metrics. It is worth noted that the proposed method
yields better performance compared to the handcrafted feature methods [2]–[5] and deep learning-based
approaches [7], [8]. The proposed method requires lower network parameters compared to that of [7] and [8].
In addition, the proposed method should be firstly considered while one conducts and implements the inverse
halftoning task.
Table 1. Performance comparisons in terms of PSNR score
Testing images ALF [3] MAP [2] LPA-ICI [4] GLDP [5] LLDO [6] SADCNN [8] CNN Inv [7] Proposed method
Koala 22.36 23.33 24.17 24.58 25.01 25.66 27.63 29.05
Cactus 22.99 23.95 25.04 25.4 25.55 25.63 27.69 29.07
Bear 21.82 22.63 23.14 23.66 24.17 - 26.35 27.98
Barbara 25.41 26.24 27.88 27.12 28.48 29.41 31.79 33.36
Shop 22.14 22.46 24.12 23.86 24.61 25.47 27.27 29.25
Peppers 30.92 28.25 30.7 30.92 31.07 32.29 31.44 33.52
Average 24.27 24.48 25.84 25.92 26.48 - 28.7 30.37
Table 2. Performance comparison in terms of SSIM value
Testing Images ALF [3] MAP [2] LPA-ICI [4] GLDP [5] LLDO [6] SADCNN [8] CNN Inv [7] Proposed method
Koala 0.6592 0.7412 0.7557 0.7831 0.7987 0.824 0.89 0.917
Cactus 0.6368 0.7746 0.7871 0.8083 0.8175 0.907 0.92 0.9589
Bear 0.6198 0.7746 0.724 0.766 0.7815 - 0.89 0.9322
Barbara 0.7138 0.7804 0.8293 0.7993 0.8463 0.882 0.92 0.952
Shop 0.64 0.6919 0.7719 0.7535 0.7976 0.866 0.89 0.9419
Peppers 0.8674 0.7681 0.8735 0.8695 0.8698 0.982 0.89 0.9865
Average 0.6895 0.7461 0.7903 0.7966 0.8185 - 0.9 0.9481
5. CONCLUSION
A simple technique for performing the inverse halftoning task has been proposed. The presented
technique exploits the usability of residual network framework in order to recover the continuous-tone image
from its halftone version. The presented method consists of a series of convolutional operators formed in the
residual block to iteratively adjust the quality of reconstructed image. The experimental section shows that
the proposed method gives the best performances compared to that of the former schemes under subjective
and objective image quality assessments. Furthermore, the deep learning apporach with residual learning can
be a good candidate for inverse halftoning technique. It is simple approach with outstanding performance.
For the future works, the proposed scheme can be extended for the other halftoning techniques, not
only for the floyd-steinberd based image halftoning. The residual framework can be simply replaced with more
sophisticated learnings such as shared source residual learning, xception module, and others. In addition,
the generative adversarial networks (GAN), transformer, linear time transformer, and recent deep learning
networks can be investigated to replace the convolutional neural networks (CNN)-based module in the
proposed method. The other loss functions such as SSIM or patch-based loss computation can also examined
in order to improve the proposed method performance. The combination of several loss functions may
improve the proposed method performance accordingly.
REFERENCES
[1] K. -L. Chung et al., “A gradient-based adaptive error diffusion method with edge enhancement,” Expert Systems with
Applications, vol. 38, no. 3, pp. 1591–1601, 2011, doi: 10.1016/j.eswa.2010.07.079.
[2] R. L. Stevenson, “Inverse halftoning via MAP estimation,” IEEE Transactions on Image Processing, vol. 6, no. 4, pp. 574–583,
1997, doi: 10.1109/83.563322.

 ISSN: 1693-6930
1334
[3] T. D. Kite, N. D. -Venkata, B. L. Evans, and A. C. Bovik, “A fast, high-quality inverse halftoning algorithm for error diffused
halftones,” IEEE Transactions on Image Processing, vol. 9, no. 9, pp. 1583–1592, 2000, doi: 10.1109/83.862639.
[4] A. Foi, V. Katkovnik, K. Egiazarian, and J. Astola, “Inverse halftoning based on the anisotropic LPA-ICI deconvolution,”
Proceedings of The 2004 International TICSP Workshop on Spectral Methods and Multirate Signal Processing, SMMSP., 2004.
[Online]. Available: https://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.80.6964&rep=rep1&type=pdf
[5] C. -H. Son, “Inverse halftoning based on sparse representation,” Optics Letters, vol. 37, no. 12, pp. 2352-2354, 2012,
doi: 10.1364/OL.37.002352.
[6] C. -H. Son and H. Choo, “Local learned dictionaries optimized to edge orientation for inverse halftoning,” IEEE Transactions on
Image Processing, vol. 23, no. 6, pp. 2542–2556, 2014, doi: 10.1109/TIP.2014.2319732.
[7] X. Hou and G. Qiu, “Image companding and inverse halftoning using deep convolutional neural networks,” arXiv preprint
arXiv:1707.00116, 2017. [Online]. Available: https://arxiv.org/pdf/1707.00116.pdf
[8] C. -H. Son, “Inverse halftoning through structure-aware deep convolutional neural networks,” Signal Processing, vol. 173, 2020,
doi: 10.1016/j.sigpro.2020.107591.
[9] H. P. A. Wicaksono, H. Prasetyo, and J. -M. Guo, “Deep learning based inverse halftoning via stationary wavelet domain,” in
2020 27th International Conference on Telecommunications (ICT), 2020, pp. 1–5, doi: 10.1109/ICT49546.2020.9239532.
[10] A. Palvanov and Y. I. Cho, “Comparisons of deep learning algorithms for MNIST in real-time environment,” International Journal of
Fuzzy Logic and Intelligent Systems, vol. 18, no. 2, pp. 126–134, 2018, doi: 10.5391/IJFIS.2018.18.2.126.
[11] S. -M. Cho and B. -J. Choi, “CNN-based recognition algorithm for four classes of roads,” International Journal of Fuzzy Logic
and Intelligent Systems 2020, vol. 20, no. 2, pp. 114–118, 2020, doi: 10.5391/IJFIS.2020.20.2.114.
[12] H. Prasetyo, A. W. H. Prayuda, C. -H. Hsia, and J. -M. Guo, “Deep concatenated residual networks for improving quality of
halftoning-based BTC decoded image,” Journal of Imaging, vol. 7, no. 2, 2021, doi: 10.3390/jimaging7020013.
[13] M. F. I. Amal, H. P. A. Wicaksono, and H. Prasetyo, “Deep residual networks for impulsive noise suppression,” in 2020 27th
International Conference on Telecommunications (ICT), 2020, pp. 1–5, doi: 10.1109/ICT49546.2020.9239522.
[14] S. -H. Jung and Y.-J. Chung, “Sound event detection using deep neural networks,” TELKOMNIKA Telecommunication,
Computing, Electronics and Control, vol. 18, no. 5, pp. 2587-2596, doi: 10.12928/telkomnika.v18i5.14246.
[15] O. Sudana, I. W. Gunaya, and I. K. G. D. Putra, “Handwriting identification using deep convolutional neural network method,”
TELKOMNIKA (Telecommunication Computing Electronics and Control), vol. 18, no. 4, pp. 1934-1941, 2020,
doi: 10.12928/telkomnika.v18i4.14864.
[16] I. B. K. Sudiatmika, F. Rahman, Trisno, and Suyoto, “Image forgery detection using error level analysis and deep learning,”
[17] J. -M. Guo, H. Prasetyo, and K. Wong, “Halftoning-based block truncation coding image restoration,” Journal of Visual
Communication and Image Representation, vol. 35, pp. 193–197, 2016, doi: 10.1016/j.jvcir.2015.12.016.
[18] H. A. Al-Jubouri and S. M. Mahmmod, “A comparative analysis of automatic deep neural networks for image retrieval,”
[19] J. -M. Guo, H. Prasetyo, and J. -H. Chen, “Content-based image retrieval using error diffusion block truncation coding features,”
IEEE Transactions on Circuits and Systems for Video Technology, vol. 25, no. 3, pp. 466–481, 2015,
doi: 10.1109/TCSVT.2014.2358011.
[20] H. Prasetyo, C. -H. Hsia, and B. A. P. Akardihas, “Halftoning-based BTC image reconstruction using patch processing with
border constraint,” TELKOMNIKA (Telecommunication Computing Electronics and Control), vol. 18, no. 1, pp. 394-406, 2020,
[21] K. He, X. Zhang, S. Ren, and J. Sun, “Identity mappings in deep residual networks,” 2016 European Conference on Computer
Vision, 2016, vol. 9908, pp. 630-645, doi: 10.1007/978-3-319-46493-0_38.
[22] A. Zaeemzadeh, N. Rahnavard, and M. Shah, “Norm-preservation: why residual networks can become extremely deep?,” in IEEE
Transactions on Pattern Analysis and Machine Intelligence, vol. 43, no. 11, pp. 3980-3990, 2021,
doi: 10.1109/TPAMI.2020.2990339.
[23] E. Agustsson and R. Timofte, “NTIRE 2017 challenge on single image super-resolution: dataset and study,” in 2017 IEEE Conference
on Computer Vision and Pattern Recognition Workshops (CVPRW), 2017, pp. 1122–1131, doi: 10.1109/CVPRW.2017.150.
[24] D. P. Kingma and J. Ba, “Adam: a method for stochastic optimization,” arXiv preprint arXiv:1412.6980, 2014. [Online].
Available: https://arxiv.org/pdf/1412.6980.pdf
[25] K. He, X. Zhang, S. Ren, and J. Sun, “Delving deep into rectifiers: surpassing human-level performance on imagenet
classification,” Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2015, pp. 1026-1034. [Online].
Available: https://openaccess.thecvf.com/content_iccv_2015/html/He_Delving_Deep_into_ICCV_2015_paper.html
BIOGRAPHIES OF AUTHORS
Heri Prasetyo received his Bachelor’s Degree from the Department of
Informatics Engineering, Institut Teknologi Sepuluh Nopember (ITS), Indonesia in 2006. He
received his Master’s and Doctoral Degrees from the Department of Computer Science and
Information Engineering, and Department of Electrical Engineering, respectively, at the
National Taiwan University of Science and Technology (NTUST), Taiwan, in 2009 and
2015, respectively. He received the Best Dissertation Award from the Taiwan Association
for Consumer Electronics (TACE) in 2015, Best Paper Awards from the International
Symposium on Electronics and Smart Devices 2017 (ISESD 2017), ISESD 2019, and
International Conference on Science in Information Technology (ICSITech 2019), and
theOutstanding Faculty Award 2019 from his current affiliated university. His research
interests include multimedia signal processing, computational intelligence, pattern
recognition, and machine learning. He can be contated at email:
heri.prasetyo@staff.uns.ac.id.

1335
Muhamad Aditya Putra Anugrah received his Bachelor’s Degree from the
Department of Informatics, Universitas Sebelas Maret (UNS), Indonesia in 2021. His
research interests include digital signal processing, image processing, deep learning, and
inverse imaging problem. He can be contated at email: mapaditya@student.uns.ac.id.
Alim Wicaksono Hari Prayuda received the bachelor degree from the
Universitas Sebelas Maret, Surakarata, Indonesia, in 2020. He is currently working toward
the master degree in Department of Electrical Engineering at National Taiwan University of
Science and Technology, Taipei, Taiwan. His research interests include Deep Learning
application for Digital Image Processing and Multimedia Engineering. He can be contated at
email: wicayudha.wy@gmail.com.
Chih-Hsien Hsia received the Ph.D. degree in Electrical Engineering from
Tamkang University, New Taipei, Taiwan, in 2010. He was an Associate Professor with the
Chinese Culture University and National Ilan University from 2015 to 2017. He is currently
a Professor with the Department of Computer Science and Information Engineering,
National Ilan University, Taiwan. Dr. Hsia is the Chapter Chair of IEEE Young
Professionals Group, Taipei Section, and the Director of the IET Taipei Local Network. He
serves as Associate Editor of the Journal of Imaging Science and Technology and the Journal
of Computers. His research interests include DSP IC Design, Multimedia Signal Processing,
and Cognitive Learning. He can be contated at email: chhsia625@gmail.com.

Deep residual neural networks for inverse halftoning

More Related Content

Similar to Deep residual neural networks for inverse halftoning (20)

More from TELKOMNIKA JOURNAL (20)

Recently uploaded (20)

Deep residual neural networks for inverse halftoning