Hindawi

Mathematical Problems in Engineering

Volume 2021, Article ID 6671613, 8 pages
https://doi.org/10.1155/2021/6671613

Review Article
Quantum Image Processing: Opportunities and Challenges

Yue Ruan , Xiling Xue , and Yuanxia Shen

School of Computer Science and Technology, Anhui University of Technology, Ma’anshan, China

Correspondence should be addressed to Yue Ruan; [email protected]

Received 20 October 2020; Revised 22 November 2020; Accepted 26 December 2020; Published 5 January 2021

Academic Editor: Wenjie Liu

Copyright © 2021 Yue Ruan et al. This is an open access article distributed under the Creative Commons Attribution License,
which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
Quantum image processing (QIP) is a research branch of quantum information and quantum computing. It studies how to take
advantage of quantum mechanics’ properties to represent images in a quantum computer and then, based on that image format,
implement various image operations. Due to the quantum parallel computing derived from quantum state superposition and
entanglement, QIP has natural advantages over classical image processing. But some related works misuse the notion of quantum
superiority and mislead the research of QIP, which leads to a big controversy. In this paper, after describing this field’s research
status, we list and analyze the doubts about QIP and argue “quantum image classification and recognition” would be the most
significant opportunity to exhibit the real quantum superiority. We present the reasons for this judgment and dwell on the
challenges for this opportunity in the era of NISQ (Noisy Intermediate-Scale Quantum).

1. Introduction classical image recognition is unknown, but it does show the

great potential of quantum image recognition.
Since the concept of quantum computing was proposed by Besides image classification and image recognition,
Feynman in 1982 [1], the achievements by many geniuses quantum image processing (QIP) involves exploiting
have shown that quantum computing has dramatically quantum properties to represent, manipulate, compress, and
improved computational efficiency. The theory to imple- address other issues related to images in a quantum com-
ment quantum computing is nearly mature; the challenge of puter. The quantum properties lead to the higher efficiency
realizing universal quantum computing mainly comes from of representing and manipulating images in the quantum
technical issues, such as manipulating large-scale qubits [2]. context but also bringing about troubles that do not exist in
In recent years, with the successional breakthroughs in the classical environment. For example, when representing
quantum technology, quantum computing has entered the images with quantum states, one cannot replicate images
era of NISQ (Noisy Intermediate-Scale Quantum), when it is since quantum states obey the noncloning principle. For
supposed as ready to display “Quantum Supremacy” in some another example, if one wants to obtain the image ma-
practical application areas [3]. nipulating result by measurements, he must find a smart way
Among these applications, quantum image processing to handle state collapsing.
(here mainly refers to the image classification and recognition) The missing or improper way to handle these issues in
is most likely to be a killer app in the near future and, for the some published papers leads to a big controversy. Some
commercial reason, to be the favorite one of big companies scholars even deem QIP is a “quantum hoax.” Honestly,
(Google, IBM, Intel, etc.). Some new works have already made their doubts make sense. Not all classical image operations
considerable progress. For example, in this February, Llyod are worthy of implementing in the quantum realm. Still, we
et al. described the method of training the map from classical note there are branch research studies in QIP that can
data to quantum states (maximizing the gap between mapped maximize quantum superiority and weaken or even elimi-
classes in Hilbert space), which has the power to distinguish the nate these troubles derived from quantum properties. We
images of ant and bee [4]. Whether the method is superior to argue “quantum image classification and recognition” would
2 Mathematical Problems in Engineering

be the candidate. We will give the reasons for this judgment Due to the merit (small storage space, simple, and easy to
after describing the research status of QIP and then discuss understand) of FRQI, many follow-up studies have been
the opportunities and challenges in this direction. carried out to extend the scheme. Instead of mapping the
pixel’s gray value to the amplitude angle, Zhang et al. used a
2. Research Status of Quantum group of ground states to represent the pixel’s value (a qubit
Image Processing string) in a larger Hilbert space. This scheme facilitates some
image operations and improves efficiency [10]. Wang et al.
The study of QIP started in 1997 [5]. In the early days, very extended FRQI to polar coordinates and replaced pixels’
few scholars paid attention to this direction, and the pub- spatial position with polar diameters and polar angles
lications were also very few. In recent years, frequent rel- [11, 12]. Ruan et al. expressed the pixel’s gray value as
evant publications in major journals indicate QIP is heating |0〉 + eiθ |1〉, replacing the cos θ|0〉 + sin θ|1〉 that stored the
up. The research topics have become rich and hierarchical, gray value in FRQI [13], which actually replaced the ZOX
which denotes QIP has become an independent research plane’s rotation on Bloch ball with a rotation of θ angle in the
branch of quantum information and quantum computing. XOY plane. All these extended schemes’ basic idea is to
In the following sections, we will discuss QIP from two prepare the image into a quantum superposition state in
aspects, quantum image format and quantum image terms of pixels’ spatial distribution information, which is not
operations. fundamentally different from FRQI.
The last quantum image format, Real Ket, was proposed
by Latorre [8]. He divided the image into 2 × 2 pieces and
2.1. Quantum Image Format. Quantum image format is the then mapped the four pixels’ grayscale value to the prob-
core topic of QIP. Qubit Lattice [6, 7], Real Ket [8], and FRQI ability amplitude of each component of a quantum state with
[9] are three major quantum image formats. 2 qubits. Equation (5) describes this quantum state, where
Qubit Lattice is the first quantum image format pro- i1 � 1 can be understood as the index of the top-left pixel,
posed by Venegas-Andraca [6, 7]. He said if the frequency i1 � 2 as the index of the top-right pixel, i1 � 3 as the bottom-
value (color value) of the light wave can be mapped to the left pixel, and i1 � 4 as the bottom-right pixel. Ci1 stores the
probability amplitude of a qubit, then the pixel value of ith mapping value of each pixel and satisfies 􏽐i1 �1,...,4|Ci1 |2 � 1.
row and the jth column can be stored in the amplitude angle 􏼌􏼌 􏼌􏼌2
􏼌􏼌 􏼌􏼌
shown in equation (1), and the whole image can be repre- 􏼌􏼌ψ 21 ×21 〉 � 􏽘 Ci1 􏼌􏼌i1 〉 st. 􏽘 􏼌􏼌􏼌Ci1 􏼌􏼌􏼌 � 1. (5)
sented as a qubit string (equation (2)). i1 �1,...,4 i1 �1,...,4
􏼌􏼌
􏼌􏼌pixel 〉 � cos θi,j |0〉 + sin θi,j |1〉, In Ref. [6], this normalized quantum state is called qudit.
􏼌 i,j
2 2 If one rewrites it in terms of qubits, a qudit is actually a
(1) superposition state of 2 qubits.
Expanding this 2 × 2 block once, one can obtain a 4 × 4
􏼌􏼌
|image〉 � 􏼚􏼌􏼌􏼌pixeli,j 〉􏼛 i � 1, 2, ..., n1 , j � 1, 2, ..., n2 . (2) block, which can be represented as a quantum state shown in
equation (6), where i2 � 1 can be regarded as the index of the
This representation scheme’s essence is to map the first (upper left corner) 2 × 2 small block, i2 � 2 is the index of
image’s spatial information to the amplitude of a single qubit the second (upper right corner) 2 × 2 small block, i2 � 3 is the
without using quantum properties of superposition and index of the third (lower left) 2 × 2 small block, and i2 � 4 is
entanglement. the index of the fourth (lower right) 2 × 2 small block. In the
The Flexible Representation of Quantum Images (FRQI) process of construction, the quantum state must be nor-
proposed by Le et al. [9] was an upgraded version of Qubit malized again, i.e., satisfies 􏽐i2 ,i1 �1,...,4|Ci2 ,i1 |2 � 1.
􏼌􏼌 􏼌 􏼌 􏼌
Lattice by exploiting quantum state superposition. The 􏼌􏼌ψ 2 2 〉 � 􏽘 Ci ,i 􏼌􏼌􏼌i2 i1 〉 st. 􏽘 􏼌􏼌􏼌􏼌Ci ,i 􏼌􏼌􏼌􏼌2 � 1.
scheme still maps each pixel’s grayscale value to the am- 2 ×2 2 1 2 1 (6)
i2 ,i1 �1,...,4 i2 ,i1 �1,...,4
plitude, meanwhile introducing an auxiliary qubit to denote
the spatial position of each pixel. Then, the whole image is Through expanding increasingly, a 2n × 2n image can be
prepared into a large quantum superposition state. Equation mapped to the quantum state as shown in the following
(4) depicts a 2n × 2n quantum image, where i can be regarded equation:
as an indicator of pixels’ position (row × column converted 􏼌􏼌 􏼌􏼌 􏼌􏼌
􏼌􏼌C
􏼌􏼌2
􏼌
|image〉 � 􏼌􏼌ψ 2n ×2n 〉 � 􏽘 Cin ,...,i1 􏼌􏼌in , ..., i1 〉 st. 􏽘 􏼌 in ,...,i1 􏼌􏼌 � 1.
to a one-dimensional vector). Due to quantum states’ su- in ,...,i1 �1,...,4 in ,...,i1 �1,...,4
perposition effect, the representation (storage) space de-
creases exponentially compared to the classical image. (7)
􏼌􏼌 The basic idea of this representation is using the basis to
􏼌􏼌pixeli 〉 � cos θi |0〉 + sin θi |1〉, (3)
represent the pixel’s spatial position while using the prob-
2n ability amplitude to represent the color information. Ref.
1 2 −1 π [13, 14] exploit this idea to implement similar schemes.
|image〉 � 􏽘 cos θi |0〉 + sin θi |1〉􏼁 ⊗ |i〉 θi ∈ 􏼔0, 􏼕.
2n i�0 2 Let us analyze the storage efficiency of these formats
(4) briefly. For a 2n × 2n gray image, if the gray value of each
Mathematical Problems in Engineering 3

pixel is represented by 8 classical bits, then a total of Local operations only involve the manipulation of a few
2n × 2n × 8 bits are required for the classical image. As for the pixels, leaving other pixels unchanged. For example, swap
Qubit Lattice, only one qubit is needed to represent a single the coordinate of ith and jth pixels. Then, there should be a
pixel’s grayscale value, so 2n × 2n qubits are needed for the unitary operator S that satisfies
whole image. In FRQI, a qubit’s gray value is still represented
by a qubit, but since the whole image is prepared into a 1 N−1􏼌􏼌
S(|image〉) � √�� 􏽘 􏼌􏼌Ck 〉 ⊗ S(|k〉),
superposition state in terms of the row and column coor- N k�0
dinates, only 2n + 1 qubits are needed. In the last format Real
Ket, as the pixels’ grayscale information is stored in the Where S(|k〉) � |k〉, k ≠ i, j and S(|i〉) � |j〉, S(|j〉) � |i〉.
probability amplitude of the components of a superposition (10)
quantum state, the storage space only need 2n qubits. Thus,
Real Ket uses minimum storage space. The operator S can be constructed as equation (11).
Swapping two pixels only needs this operator to act once.

2.2. Quantum Image Operation. Image operations involve a S � |i〉〈j| +|j〉〈i| + 􏽘 |k〉〈k|. (11)
wide variety of types. In recent years, many works have k≠i,j

discussed this topic in the quantum realm, and the opera-

Consider the counterpart operation in a classical image.
tions realized are increasingly abundant. Here, we roughly
Swapping two pixels is equivalent to switching i and j in an
classify these operations into geometric transformation,
array. That means one needs to perform at least two writes,
color transformation, and complicated operations such as
one at i and one at j. Obviously, local quantum operations
compression and retrieval. We describe them, respectively,
are also faster than classical processes.
as follows.

2.2.1. Geometric Transformation. Geometric transformation 2.2.2. Color Transformation. Color transformation refers to
refers to changing the spatial position (coordinate) of pixels, changing the image’s pixel value, such as halftone processing
such as image rotation, or changing image shape, like zoom of images [23]. If such changes can be restored by some
in or out. The realized operations include rotation, the top/ means, then image watermark [24–26], shuffling color
bottom swapping or left/right swapping, the interchange of blocks [27, 28], image encryption/decryption [29–33], etc.,
any two coordinates of the image [15–17], erosion and di- can be regarded as this category.
lation operations [18], image magnification [19, 21], and In terms of the acting scope of color transformation, one
overall translation and cyclic translation of the image [22]. can also divide it into global operation and local operation.
Take FRQI as an example. Its format can be simplified as Using a similar analysis above, one can see that the global
qubits representing the color information tensor by qubits color transformation achieves exponential acceleration
representing pixels’ coordinate information (equation (8)). k compared to the classical counterpart and local operation is
is the position component, representing N pixel’s spatial faster than its classical counterpart [10, 34]. Because the
position. Ck is the color component, representing the kth quantum Fourier transform is invertible and exponentially
pixel’s gray value. faster than the classical Fourier transform, it is widely used
in image watermarking and image encryption/decryption
1 N−1􏼌􏼌 [24, 29, 30]. Due to the “uncertainty” principle of quantum
|image〉 � √�� 􏽘 􏼌􏼌Ck 〉 ⊗ |k〉. (8) mechanics, quantum image watermarking, quantum image
N k�0
encryption/decryption, etc., can guarantee hidden infor-
In general, geometric transformation refers to realizing mation security.
the image operation operator acting on the position com-
ponent. According to the scope of the operator, we can
classify it into global operation and local operation. 2.2.3. Complicated Image Operations. In classical image
As the name implies, global operations operate on all processing, effective methods to achieve compression, re-
pixels of the image, such as rotation (rotate 90∘ , 180∘ , 270∘ ). trieval, recognition, segmentation, registration, and other
Equation (9) gives a formal description of such operations, operations generally need to perform some preprocessing
where G is the unitary transformation acting on all pixels. based on the original image data, such as transforming the
image domain or extracting image features. In this paper, we
1 N−1􏼌􏼌 call these operations as “complicated” image operations.
G(|image〉) � √�� 􏽘 􏼌􏼌Ck 〉 ⊗ G(|k〉). (9)
N k�0 Compression is the most discussed topic. In Ref. [9],
quantum image compression is defined as reducing the
Since each pixel’s gray values are stored in the corre- number of quantum gates when preparing quantum images
sponding component of the superposition state, performing (quantum states). This definition is different from classical
a unitary operation G would modify all components si- image compression, and in essence, the discussion of such
multaneously (quantum parallelism). Thus, global operation problems can be defined as the optimization of quantum
G’s efficiency is much higher than the counterpart operation circuits. Ref. [10, 11, 34, 35] describe the methods of per-
to a classical image. forming such compression. Although these methods are
4 Mathematical Problems in Engineering

slightly different, the basic idea is to reduce the number of image manipulation result by measurements, the claim that
quantum gates by simplifying Boolean expressions. A more exponential acceleration of quantum image operation may
natural definition of compression would be how to reduce not exist. Mastriani enumerated these doubts [42] (similar
the number of qubits representing a quantum image (cor- challenges in quantum machine learning [43]) and con-
responding to classical image compression), that is, to re- cluded that many published works related to QIP are
duce the dimensions of the Hilbert space representing the “Quantum Hoax.” His main viewpoints can be summarized
quantum state (quantum image). From the perspective of as follows:
information theory, it seeks to represent a quantum image
(1) Input: many published papers did not consider/
with more concentrated energy, i.e., concentrate image
reckon the cost of preparing quantum states (images)
primary information into a smaller dimension. Ref. [8]
from classical data (image).
proposed a transforming method based on the matrix
product state theory. And this mathematical form makes it (2) Output: obtaining the result of image operation
possible to obtain the optimum lossless compression and requires an exponential scale of measurements.
enables us to distinguish the important information and (3) Noise: quantum image is sensitive to noise and
redundant information to perform lossy compression. simulation software such as MATLAB is not capable
Another frequently discussed topic is image retrieval. In of verifying the correctness of quantum algorithms.
quantum image processing, retrieval has two meanings: one
His criticism triggered a fierce debate. This March, the
is to retrieve classical information from a quantum state and
journal Quantum Information Processing published both Li
the other is to retrieve the quantum information in the
et al.’s comments [44] on Mastriani’s original paper and
quantum state. The former’s implemental method has been
Mastriani’s rebuttal to these comments [45]. The opinions
described in Ref. [5, 6, 35–37]. With the same idea, a large
from both sides partly make sense, but neither seems to be
number of quantum images (quantum states) are prepared,
quite right.
the probability amplitude of each ground state is estimated
For the first doubt, FRQI or other image formats are just
by measurements repeatedly, and then the original quantum
a method to map classical data to quantum data without
image (quantum states) is recovered according to the
theoretical defects. These quantum images can be efficiently
probability distribution. The latter is based on the retrieval of
prepared. The interested reader can refer to Yao et al.’s work
quantum image content. Schutzhold presented a quantum
[14] for a detailed discussion on this issue.
algorithm for finding simple patterns (such as a parallel line)
For the second and third doubts, if image operations like
in black-and-white binary images [38]. The algorithm uti-
geometric transformation and color transformation are
lizes quantum Fourier transform characteristics to work in
one’s final goals, he has to measure each pixel to get the
parallel and can achieve exponential speedup compared to
results. That procedure for a quantum state (quantum im-
classical algorithms. Venegas-Andraca described the rela-
age) is named quantum state tomography, which requires an
tionships between the vertices of graphs such as triangles and
exponential scale of measurements for a general state [46].
squares with quantum entanglement and exploited Bell
Besides, since the measurement result is the statistical result
inequality to provide a method to retrieve the existence of
of the observed value, it is difficult to eliminate the mea-
these graphs in black-and-white binary images [39].
surement noise. Thus, in quantum image processing, this
Besides, Le et al. discussed image segmentation. The
kind of research work would have little practical significance.
algorithm uses the operator prepared by an orthogonal basis
But, if one only wants to take advantage of the quantum
and the gray level information encoded into the ground state
image’s overall characteristics or some statistical properties
of the quantum state (orthogonal basis) to make an
rather than read all pixels’ value. Quantum image operations
equivalent determination and uses Grover algorithm to
like geometric transform and color transform are inter-
accelerate this process [40]. Caraiman realized image seg-
mediate steps to the end. These operations would make
mentation based on the threshold by calculating histogram
sense. For example, the HHL algorithm [47] for solving
[41]. Zhang et al. discussed image registration by giving an
linear equations presents the solution hidden in the quan-
ordinal number to images with different rotation angles and
tum superposition state. The correct way to use this solution
then using the Grover algorithm to retrieve the ordinal
is to exploit this state’s overall property rather than measure
number [12]. This approach is similar to binding a keyword
it to get the probability distribution for each component.
to each image and then retrieving it based on the keyword,
This approach to using HHL forms the foundation of many
rather than content-based retrieval.
quantum machine learning algorithms [48–50].
Whether the simulation software MATLAB can verify
2.3. Discussion. Applying the quantum properties of su- the algorithm of quantum image processing should be
perposition and entanglement to map classical images and considered from two aspects. First, quantum image opera-
store them in qubits is the basic idea of preparing quantum tions are certain quantum algorithms which can be de-
images. Due to the parallel computing induced by the scribed by sequential unitary transformations acting on a
quantum superposition effect, the quantum image opera- complex vector in a Hilbert space. Certainly, MATLAB has
tion’s efficiency is much higher than the corresponding the capability to simulate such unitary evolution. Second, to
classical image operation. But if taking into account the cost our knowledge, there seems to be no module in MATLAB
of quantum image preparation and the cost of obtaining the that can simulate quantum noise. Therefore, if one wants to
Mathematical Problems in Engineering 5

verify quantum algorithms’ performance on real quantum The circuit in Figure 1(b), proposed by Havlı́ček et al. in
computing devices, IBM’s qiskit or other tools would be a 2019, is known as the “inversion test” to calculate the dis-
better choice. tance between quantum states. x and y are features of the
classical data (image), as an argument to input into quantum
3. Opportunities and Challenges circuit Φ and its inverse circuit Φ† . One can see that if x and
y are equal (similar), the probability should eventually be 1
3.1. Opportunities. Quantum algorithms can improve or near to be 1 when doing the projection measurement on
computational efficiency, but it may not apply to all ap- |0 . . . 00〉. If x and y differ a lot, this probability is smaller.
plication scenarios. The successful quantum algorithms all This technique can also calculate distances in quantum space
have a distinct characteristic: the intermediate procedure of (Hilbert space), and its advantage versus “swap test” is re-
computation is complicated, suitable for constructing ducing a half number of qubits used.
quantum state entanglement and superposition for parallel
acceleration; the result is simple and often a decisive answer.
3.2. Challenges. Based on the above discussion, one could
In that case, entanglement/superposition degenerates at the
conclude that quantum image classification and recognition
end of the quantum algorithm, and the probability ampli-
is the most significant opportunity in quantum image
tude of a single basis is 1 or close to 1. For quantum image
processing. But we must note that the certification of its
processing, the task of image classification and recognition
ability beyond the classical image recognition method in
accords with this characteristic. The intermediate algorithm
both theory and practice still needs to face the following
execution procedures involve feature extraction, classifier
challenges.
training, and various distance computing. The result only
needs to answer “yes or no” to determine whether the image
belongs to a specific category. 3.2.1. Preparation of Quantum Images. Only by mapping
Moreover, a wide variety of machine learning protocols and preparing classical data to quantum superposition state
operate by performing matrix operations on vectors in a can the advantage of quantum computation be brought into
high-dimensional vector space. Quantum mechanics is all play. The general method to prepare a quantum superpo-
about matrix operations on vectors in high-dimensional sition state is to use QRAM (Quantum Random Access
vector spaces. Thus, performing machine learning tasks in Memory) [54, 55]. Its basic idea is to use a “bucket brigade”
the quantum realm would probably be beneficial if one can structure to distribute N d-dimensional vector data on the
take advantage of the natural connection between these two Nd leaf nodes of the “tree.” Based on this structure, QRAM
disciplines. This path has proved to be correct. Quantum can prepare N d-dimensional vectors into log(Nd) qubits
machine learning has made a lot of achievements in recent superposition state in O(log(Nd)) time. However, this
years. Thus, quantum machine learning algorithms could act structure requires O(Nd) physical resources. Its scale is
as the building blocks for the classification and recognition exponential in the number of qubits, so whether it can
of quantum images. Here, we would like to introduce two provide real computational advantages in the actual ex-
commonly used techniques “swap test” and “inversion test” perimental environment is still a big question [56, 57].
to show the advantages of these methods.
Figure 1(a) illustrates the principle of “swap test.” |x〉
3.2.2. Feature Extraction. Feature extraction plays a vital
and |y〉 are two quantum states (quantum images). Using
role in classical image recognition. Effective features can
the circuit, one can estimate the similarity between |x〉 and
eliminate the influence of image background, size, lighting
|y〉 by measuring whether auxiliary qubit ends up in |0〉
conditions, camera angle, etc., and improve image recog-
alone.
nition accuracy; however, little literature talks about how to
􏼌􏼌 1 1 extract features from the mainstream quantum image for-
P􏼐􏼌􏼌0anc 〉􏼑 � + |〈x | y〉|2 . (12)
2 2 mats. The absence of discussion on this issue cast doubt on
some impressive work. For example, Ref. [14] used Hada-
|〈x | y〉| is known as “fidelity” in quantum information
mard transform to calculate the difference between adjacent
and “cosine distance” in classical machine learning. One can
pixels quantum image and then compared the known
see that if |x〉 and |y〉 are farthest away from each other
pattern by “swap test” to detect the image edge. However, if
(orthogonal), this probability is 1/2; if |x〉 and |y〉 are closest
the image pattern has a certain degree of deformation, such
to each other, this probability is 1. One can also see that this
as size inconsistency, which will lead to a considerable
probability estimation has nothing to do with the feature
difference between the known pattern and the quantum
space’s dimension. The higher the dimension is, the more
image to be tested, is it still valid to use “swap test” to detect
efficient this technique is against the classical algorithm.
the image edge? Moreover, this paper only experiments for
Lloyd points out that even considering the cost of preparing
image edge detection on binary images; it is unknown that
quantum states, this technique is much more efficient than
the proposed approach for natural images’ detection is still
distance calculations on classical computers [51]. Derived
effective.
from this technique, many more general distances such as
􏽰����������
Euclidean distance � 2 − 2|〈x|y〉| can also be calculated.
This technique is widely used in various quantum machine 3.2.3. Nonlinear Operations. Effective procedures in a
learning algorithms [49–53]. classical image (pattern) recognition and machine learning
6 Mathematical Problems in Engineering

0 H H 0···0 Ф(x) Ф(y)†

x
y
(a) (b)

Figure 1: Quantum computing techniques for distance estimation. (a) Swap test. (b) Inversion test.

tend to be nonlinear, such as the sigmoid function used by We are sure that not all classical image manipulation is
training perceptron in deep learning. The nature of quantum necessary to implement in a quantum computer. Whether
mechanics is linear. A feasible scheme is to induce nonlinear image geometric transformation, color transformation, and
operation through measurements. However, the quantum other similar operations are worthy of implementing in the
state will collapse after measurements, making the system’s quantum realm depends on the application scenario. For
evolution lose its quantum characteristics and degenerate example, Yao et al. calculated the difference between adja-
into the classical probability perceptron. Another popular cent pixels in the quantum image via Hadamard transform
solution is to implement nonlinear operation (such as (which can be regarded as a kind of color transform) and
feature extraction) by a classical process and then compute then used the result directly to obtain the image edge by
the kernel function in quantum space (Hilbert space). Llyod matching a known pattern [14]. That is, these kinds of
called this approach as “quantum embedding” [4, 52, 58]. quantum image operations make sense, only proving that
This strategy’s essence is to circumvent the nonlinear op- they are intermediate steps to the end rather than the final
eration in quantum space; whether it is superior to the goal.
classical machine learning algorithm needs further Whether quantum image processing (in some aspects or
verification. some specific applications) can achieve an advantage over
classical image processing in realistic scenes remains to be
seen. It depends on the solution of four challenges in Section
3.2.4. Noise. Noise is one of the most important theoretical 3.2, some of which are unique to the quantum image
problems in quantum computation. The existence of processing, such as the quantum image feature extraction;
quantum noise may lead to quantum computation to some are common to quantum algorithms, such as handling
classical probabilistic computation. Since the 1990s, people noise. Figuring out one will significantly boost the research
have been studying quantum error correction codes [59] and of this field.
further developing the concept of fault-tolerant quantum In short, quantum image processing with both oppor-
computing [60]. In recent years, eliminating errors caused tunities and challenges is worthy of further in-depth study.
by noise has been a fascinating research direction, among
which topological quantum computing has aroused the most
concern because it promises to solve quantum noise com- Data Availability
pletely [61]. The more interesting thing is that functions that The data used to support the findings of this study are in-
cannot be learned in a noisy classical environment can be cluded within the article.
learned in the noisy quantum environment, such as Dis-
junctive Normal Form (DNF) [62]. In recent years, this
research has extended to more general linear functions such Conflicts of Interest
as odd and even functions [63]. The authors declare that there are no conflicts of interest
regarding the publication of this paper.
4. Conclusions
Compared with classical image processing, quantum image Acknowledgments
processing is far from sufficient in both depth and width. The This research was funded by the Natural Science Founda-
“quantum advantage” claimed in some related published tion, China (grant no. 61802002), Natural Science Foun-
papers has also been doubted by many scholars; the core of dation of Anhui Province, China (grant no. 1708085MF162),
these doubts is “how to obtain quantum image operating and Foundation for Key Project of Anhui Provincial De-
results efficiently and accurately.” We deem that these partment of Education (grant nos. KJ2019A0063 and
studies try to get quantum image operation results by re- KJ2020A0233).
covering classical images via measurements without prac-
tical significance.
On the contrary, the kind of research that only exploits References
quantum images’ statistical characteristics may be valuable. [1] R. P. Feynman, “Simulating physics with computers,” Inter-
With the assistance of quantum machine learning, quantum national Journal of Theoretical Physics, vol. 21, no. 6-7,
image classification and recognition would have the most pp. 467–488, 1982.
significant opportunity to be a “killer app” in the actual [2] C. R. Chang, “The second quantum revolution with quantum
commercial field in the NISQ era. computers,” AAPPS Bulletin, vol. 30, no. 1, pp. 9–22, 2020.
Mathematical Problems in Engineering 7

[3] J. Preskill, “Quantum computing in the NISQ era and be- ratio,” Quantum Information Processing, vol. 14, no. 11,
yond,” Quantum, vol. 2, p. 79, 2018. pp. 4001–4026, 2015.
[4] S. Lloyd, M. Schuld, A. Ijaz, J. Izaac, and N. Killoran, [22] J. Wang, N. Jiang, and L. Wang, “Quantum image transla-
“Quantum embeddings for machine learning,” 2020, https:// tion,” Quantum Information Processing, vol. 14, no. 5,
arxiv.org/abs/2001.03622. pp. 1589–1604, 2015.
[5] A. Y. Vlasov, “Quantum computations and images recogni- [23] C.-C. Tseng and T.-M. Hwang, “Quantum digital image
tion,” 1997, https://arxiv.org/abs/quant-ph/9703010. processing algorithms,” in Proceedings of 16th IPPR Confer-
[6] S. Venegas-Andraca and S. Bose, “Storing, processing, and ence on Computer Vision, Graphics and Image Processing,
retrieving an image using quantum mechanics,” in Proceed- Kinmen, China, August 2003.
ings of Quantum Information and Computation, vol. 5105, [24] W.-W. Zhang, F. Gao, B. Liu, H.-Y. Jia, Q.-Y. Wen, and
International Society for Optics and Photonics, Orlando, FL, H. Chen, “A quantum watermark protocol,” International
USA, August 2003. Journal of Theoretical Physics, vol. 52, no. 2, pp. 504–513, 2013.
[7] S. Venegas-Andraca and S. Elı́as, “Discrete quantum walks [25] W.-W. Zhang, F. Gao, B. Liu, Q.-Y. Wen, and H. Chen, “A
and quantum image processing,” Ph.D. thesis, University of watermark strategy for quantum images based on quantum
Oxford, Oxford, UK, 2005. Fourier transform,” Quantum Information Processing, vol. 12,
[8] J. I. Latorre, “Image compression and entanglement,” 2005, no. 2, pp. 793–803, 2013.
https://arxiv.org/abs/quant-ph/0510031. [26] F. Yan, A. M. Iliyasu, B. Sun, S. E. Venegas-Andraca, F. Dong,
[9] P. Q. Le, Q. Phuc, K. F. Hirota, and K. Hirota, “A flexible and K. Hirota, “A duple watermarking strategy for multi-
representation of quantum images for polynomial prepara- channel quantum images,” Quantum Information Processing,
tion, image compression, and processing operations,” vol. 14, no. 5, pp. 1675–1692, 2015.
Quantum Information Processing, vol. 10, no. 1, pp. 63–84, [27] N. Jiang, W.-Y. Wu, and L. Wang, “The quantum realization
2011. of Arnold and Fibonacci image scrambling,” Quantum In-
[10] Y. Zhang, K. Lu, Y. Gao, and M. Wang, “NEQR: a novel formation Processing, vol. 13, no. 5, pp. 1223–1236, 2014.
enhanced quantum representation of digital images,” [28] R.-G. Zhou, Y.-J. Sun, and P. Fan, “Quantum image Gray-
Quantum Information Processing, vol. 12, no. 8, pp. 2833– code and bit-plane scrambling,” Quantum Information Pro-
2860, 2013. cessing, vol. 14, no. 5, pp. 1717–1734, 2015.
[11] M. Wang, K. Lu, Y. Zhang, and X. Wang, “FLPI: represen- [29] R.-G. Zhou, Q. Wu, M.-Q. Zhang, and C.-Y. Shen, “Quantum
tation of quantum images for log-polar coordinate,” in Pro- image encryption and decryption algorithms based on
ceedings of 5th International Conference on Digital Image quantum image geometric transformations,” International
Processing (ICDIP 2013), vol. 8878, International Society for Journal of Theoretical Physics, vol. 52, no. 6, pp. 1802–1817,
Optics and Photonics, Beijing, China, July 2013. 2013.
[12] Y. Zhang, K. Lu, Y. Gao, and K. Xu, “A novel quantum [30] Y.-G. Yang, X. Jia, S.-J. Sun, and Q.-X. Pan, “Quantum
representation for log-polar images,” Quantum Information cryptographic algorithm for color images using quantum
Processing, vol. 12, no. 9, pp. 3103–3126, 2013. Fourier transform and double random-phase encoding,”
[13] Y. Ruan, H. Chen, Z. Liu, and J. Tan, “Quantum image with Information Sciences, vol. 277, pp. 445–457, 2014.
high retrieval performance,” Quantum Information Process- [31] X.-H. Song and X.-M. Niu, “Comment on: novel image en-
ing, vol. 15, no. 2, pp. 637–650, 2016. cryption/decryption based on quantum Fourier transform
[14] X.-W. Yao, H. Wang, Z. Liao et al., “Quantum image pro- and double phase encoding,” Quantum Information Pro-
cessing and its application to edge detection: theory and cessing, vol. 13, no. 6, pp. 1301–1304, 2014.
experiment,” Physical Review X, vol. 7, no. 3, Article ID [32] X.-H. Song, S. Wang, A. A. Abd El-Latif, and X.-M. Niu,
031041, 2017. “Quantum image encryption based on restricted geometric
[15] Y.-G. Yang, J. Xia, X. Jia, and H. Zhang, “Novel image en- and color transformations,” Quantum Information Processing,
cryption/decryption based on quantum Fourier transform vol. 13, no. 8, pp. 1765–1787, 2014.
and double phase encoding,” Quantum Information Pro- [33] N. R. Zhou, T. X. Hua, L. H. Gong, D. J. Pei, and Q. H. Liao,
cessing, vol. 12, no. 11, pp. 3477–3493, 2013. “Quantum image encryption based on generalized Arnold
[16] P. Q. Le, A. M. Iliyasu, and F. Dong, “Fast geometric transform and double random-phase encoding,” Quantum
transformations on quantum images,” International Journal Information Processing, vol. 14, no. 4, pp. 1193–1213, 2015.
of Applied Mathematics, vol. 40, no. 3, pp. 113–123, 2010. [34] H.-S. Li, Z. Qingxin, S. Lan, C.-Y. Shen, R. Zhou, and J. Mo,
[17] P. Q. Le, A. M. Iliyasu, F. Dong, and K. Hirota, “Strategies for “Image storage, retrieval, compression and segmentation in a
designing geometric transformations on quantum images,” quantum system,” Quantum Information Processing, vol. 12,
Theoretical Computer Science, vol. 412, no. 15, pp. 1406–1418, no. 6, pp. 2269–2290, 2013.
2011. [35] H.-S. Li, Q. Zhu, R.-G. zhou, M.-C. Li, l. Song, and H. Ian,
[18] S. Iliyasu, X. Mao, L. Chen, and Y. Xue, “Quantum digital “Multidimensional color image storage, retrieval, and com-
image processing algorithms based on quantum measure- pression based on quantum amplitudes and phases,” Infor-
ment,” Optik, vol. 124, no. 23, pp. 6386–6390, 2013. mation Sciences, vol. 273, pp. 212–232, 2014.
[19] R.-G. Zhou, Z.-b. Chang, P. Fan, W. Li, and T.-t. Huan, [36] G. Beach, C. Lomont, and C. Cohen, “Quantum image
“Quantum image morphology processing based on quantum processing (quip),” in Proceedings of 32nd Applied Imagery
set operation,” International Journal of Theoretical Physics, Pattern Recognition Workshop, IEEE, Wasington, DC, USA,
vol. 54, no. 6, pp. 1974–1986, 2015. October 2003.
[20] N. Jiang and L. Wang, “Quantum image scaling using nearest [37] C. Y. Pang, W. Z. Zheng, and G. C. Guo, “Quantum discrete
neighbor interpolation,” Quantum Information Processing, cosine transform for image compression,” 2006, https://arxiv.
vol. 14, no. 5, pp. 1559–1571, 2015. org/abs/quant-ph/0601043.
[21] N. Jiang, J. Wang, and Y. Mu, “Quantum image scaling up [38] R. Schützhold, “Pattern recognition on a quantum computer,”
based on nearest-neighbor interpolation with integer scaling Physical Review A, vol. 67, no. 6, Article ID 062311, 2003.
8 Mathematical Problems in Engineering

[39] S. E. Venegas-Andraca and J. L. Ball, “Processing images in [61] D. Aasen, M. Hell, R. V. Mishmash et al., “Milestones toward
entangled quantum systems,” Quantum Information Pro- Majorana-based quantum computing,” Physical Review X,
cessing, vol. 9, no. 1, pp. 1–11, 2010. vol. 6, no. 3, Article ID 031016, 2016.
[40] P. Q. Le, A. M. Iliyasu, A. M. Iliyasu, F. Dong, and K. Hirota, [62] N. H. Bshouty and J. C. Jackson, “Learning DNF over the
“Efficient color transformations on quantum images,” Journal uniform distribution using a quantum example oracle,” SIAM
of Advanced Computational Intelligence and Intelligent In- Journal on Computing, vol. 28, no. 3, pp. 1136–1153, 1998.
formatics, vol. 15, no. 6, pp. 698–706, 2011. [63] A. B. Grilo, I. Kerenidis, and T. Zijlstra, “Learning-with-errors
[41] S. Caraiman and V. I. Manta, “Histogram-based segmentation problem is easy with quantum samples,” Physical Review A,
of quantum images,” Theoretical Computer Science, vol. 529, vol. 99, no. 3, Article ID 032314, 2019.
pp. 46–60, 2014.
[42] M. Mastriani, “Quantum image processing?” Quantum In-
formation Processing, vol. 16, no. 1, p. 27, 2017.
[43] J. Biamonte, P. Wittek, N. Pancotti, P. Rebentrost, N. Wiebe,
and S. Lloyd, “Quantum machine learning,” Nature, vol. 549,
no. 7671, pp. 195–202, 2017.
[44] H.-S. Li, P. Fan, H. Y. Xia, and R. G. Zhou, “A comment on
“quantum image processing?”” Quantum Information Pro-
cessing, vol. 19, no. 5, p. 155, 2020.
[45] M. Mastriani, “Quantum image processing: the pros and cons
of the techniques for the internal representation of the image.
A reply to: a comment on “quantum image processing?””
Quantum Information Processing, vol. 19, no. 5, p. 156, 2020.
[46] Quantum Tomography. http://en.wikipedia.org/wiki/Quantum
tomography.
[47] A. W. Harrow, A. Hassidim, and L. Seth, “Quantum algorithm
for linear systems of equations,” Physical Review Letters,
vol. 103, no. 15, Article ID 150502, 2009.
[48] S. Lloyd, M. Mohseni, and P. Rebentrost, “Quantum principal
component analysis,” Nature Physics, vol. 10, no. 9,
pp. 631–633, 2014.
[49] P. Rebentrost, M. Mohseni, and S. Lloyd, “Quantum support
vector machine for big data classification,” Physical Review
Letters, vol. 113, no. 13, Article ID 130503, 2014.
[50] N. Wiebe, D. Braun, and S. Lloyd, “Quantum algorithm for
data fitting,” Physical Review Letters, vol. 109, no. 5, Article ID
050505, 2012.
[51] S. Lloyd, M. Mohseni, and P. Rebentrost, “Quantum algo-
rithms for supervised and unsupervised machine learning,”
2013, https://arxiv.org/abs/1307.0411.
[52] V. Havlı́ček, A. D. Córcoles, K. Temme et al., “Supervised
learning with quantum-enhanced feature spaces,” Nature,
vol. 567, no. 7747, pp. 209–212, 2019.
[53] N. Wiebe, A. Kapoor, and K. Svore, “Quantum nearest-
neighbor algorithms for machine learning,” Quantum In-
formation and Computation, vol. 15, pp. 318–358, 2018.
[54] V. Giovannetti, S. Lloyd, and L. Maccone, “Architectures for a
quantum random access memory,” Physical Review A, vol. 78,
no. 5, Article ID 052310, 2008.
[55] V. Giovannetti, S. Lloyd, and L. Maccone, “Quantum random
access memory,” Physical Review Letters, vol. 100, no. 16,
Article ID 160501, 2008.
[56] S. Aaronson, “Read the fine print,” Nature Physics, vol. 11,
no. 4, pp. 291–293, 2015.
[57] J. Adcock, E. Allen, M. Day et al., “Advances in quantum
machine learning,” 2015, https://arxiv.org/abs/1512.02900.
[58] M. Schuld and K. Nathan, “Quantum machine learning in
feature Hilbert spaces,” Physical Review Letters, vol. 122, no. 4,
Article ID 040504, 2019.
[59] E. Knill, R. Laflamme, and L. Viola, “Theory of quantum error
correction for general noise,” Physical Review Letters, vol. 84,
no. 11, p. 2525, 2000.
[60] A. M. Steane, “Efficient fault-tolerant quantum computing,”
Nature, vol. 399, no. 6732, pp. 124–126, 1999.