Quantum Image Processing

Alok Anand Meizhong Lyu Prabh Simran Baweja

Electrical & Computer Engineering Material Sciences and Engineering Electrical & Computer Engineering
Carnegie Mellon University Carnegie Mellon University Carnegie Mellon University
Pittsburgh, PA, U.S.A. 15213 Pittsburgh, PA, U.S.A. 15213 Pittsburgh, PA, U.S.A. 15213
[email protected] [email protected] [email protected]

Vinay Patil
Electrical & Computer Engineering
arXiv:2203.01831v1 [quant-ph] 1 Mar 2022

Carnegie Mellon University

Pittsburgh, PA, U.S.A. 15213
[email protected]

Abstract—Image processing is popular in our daily life because B. Quantum Image Representation (QIR)
of the need to extract essential information from our 3D world,
including a variety of applications in widely separated fields Quantum image processing is one of the most attractive and
like bio-medicine, economics, entertainment, and industry. The promising tools within the Quantum Technology toolbox. The
nature of visual information, algorithm complexity, and the representation of an image on a quantum computer in the form
representation of 3D scenes in 2D spaces are all popular research of normalized states facilitates numerous image processing
topics. In particular, the rapidly increasing volume of image problems [2].
data as well as increasingly challenging computational tasks
have become important driving forces for further improving the
efficiency of image processing and analysis. Since the concept
of quantum computing was proposed by Feynman in 1982,
many achievements have shown that quantum computing has
dramatically improved computational efficiency [1]. Quantum
information processing exploit quantum mechanical properties,
such as quantum superposition, entanglement and parallelism,
and effectively accelerate many classical problems like factoring
Fig. 1. Scheme of Quantum Image Processing (QImP)
large numbers, searching an unsorted database, Boson sampling,
quantum simulation, solving linear systems of equations, and
machine learning. These unique quantum properties may also be The following methods have been proposed for QIR:
used to speed up signal and data processing. In quantum image 1) Qubit Lattice: In 2003, Venegas-Andraca and Bose
processing, quantum image representation plays a key role, which suggested the Qubit Lattice model [3] to map image’s spatial
substantively determines the kinds of processing tasks and how information with the amplitude of a single qubit without using
well they can be performed.
quantum properties, therefore requiring the same number of
Index Terms—Feature Map, QPIE, Hadamard, QSobel, Quan- qubits as pixels. This is a quantum-analog representation of
tum CNN, Edge Detection, K-means, Corner detection, Quantum classical images. The pixel value of ith row and the j th column
kernel, SVM, CIFAR, Noise Probability, Pauli Noise, Depolarising can be stored as:
noise    
θi,j θi,j
|pixeli,j i = (cos |0i + sin |1i) (1)
I. L ITERATURE R EVIEW 2 2
A. Classical Image Representation 2) Flexible Representation of Quantum Images (FRQI):
Images in classical computers are defined as matrices of FRQI, proposed by Le et al. [4], maps each pixel’s grayscale
numbers representing the discrete color or intensity values value to the amplitude as well as captures the corresponding
present in every image pixel. Each image is considered as positions in an image and integrates them into a quantum state.
input data displayable in numerous ways, whether as arrays The FRQI representation is expressed as:
of pixel values or either multidimensional plots representing
2n−1
the distribution of pixel intensities. Images can be rendered 1 X
in color layered with 3 channels (Blue, Green, and Red), |I(θ)i = (sin (θi ) |0i + cos (θi ) |1i) |ii (2)
2n i=0
Grayscale with pixel values varying from 0 (black) to 255
(white), and binary portraying black or white values (0 or 1) where θi encodes pixel value of the corresponding position
only. |i>. The FRQI representation maintains a normalized state and
the representation space decreases exponentially compared to where
the classical image due to the quantum states’ superposition T
I 0 = (I1,1 , I2,1 , . . . , Ir,1 , I1,2 , . . . , Ir,c ) (7)
effect. FRQCI and IFRQI are two improved flexible represen-
tation of quantum images reported by Li et al. [5] and Khan and
et al. [6] respectively. I 0 (i)
ci = (8)
3) Novel Enhanced Quantum Representation (NEQR): ||I 0 ||
Zhang et al. [7] reported a representation that uses the basis is the normalized value of the ith element of the vector.
state of a qubit sequence to store the grayscale value of every This method allows one to encode images of arbitrary size r
pixel instead of probability amplitude encoded in a qubit as × c However, this encoding has the problem of the extraction
in FRQI. The images are stored by entangling color infor- of the exact original image from its encoding quantum circuit.
mation, represented by |f (y, x)i, with location information, This is because the pixel value is stored in the state amplitude,
represented by |yxi. The NEQR representation for a 2n × 2n so that its approximate value can only be restored by multiple
image is expressed as: measurements.
2n 2n
6) Quantum Feature Map based Data Encoding: Classical
2 −1 2 −1
1 X X data is encoded to the quantum state space using a quantum
|Ii = |f (y, x)i|yxi (3)
2n y=0 x=0 feature map. The choice of the feature map to use is important
and depends on the dataset to be classified. A quantum feature
where f (y, x) refers to the pixel intensity at f (y, x). NEQR map utilises classical feature vector for encoding to quantum
could perform the complex and elaborate color operations state, that includes applying the unitary operation on the initial
more conveniently than FRQI does. NEQR can achieve a state.
quadratic speedup in quantum image preparation and retrieve
digital images from quantum images accurately. However,
NEQR representation uses more qubits to encode a quantum
image. Researchers have proposed a few variations for NEQR
as well: Improved NEQR (INEQR) [8], generalized model of
NEQR (GNEQR) [9] and CQIR [10].
4) Normal Arbitrary Quantum Superposition State
(NAQSS): NAQSS was proposed to solve multi-dimensional
color image processing by Li et al. [11]. The NAQSS is Fig. 2. Feature map on N-qubits generated by the unitary function from
a(n + 1)- qubit quantum representation, which can be used to Havlicek et al.’s work [13]
represent multi-dimensional color images. N qubits represent
colors and coordinates of 2n pixels and the remaining 1 qubit Feature map contain layers of Hadamard gates with entan-
represents an image segmentation information to improve the gling blocks encoded as,
accuracy of image segmentation. The NAQSS corresponds to
the color and angle of the image one to one, mapping the
color information to a certain value on the interval [0, π/2].
The NAQSS can be represented as:
n
2X −1
|Ii = θi |v1 i|v2 i. . . |vk i ⊗ |χi i (4)
i=0

where
|χi i = cos γi |0i + sin γi |1i (5)
represents the segmentation information. NAQSS can improve
the efficiency and accuracy of image segmentation. However,
it cannot accurately measure the pixels of the image.
Fig. 3. Feature map on N-qubits generated by the unitary function from
5) Quantum Probability Image Encoding (QPIE): Yao Havlicek et al.’s work [13]
et al.’s QPIE [12] representation has an immediate advantage
over the other encoding methods previously discussed is that Here, data encoding employs ZZFeature Map , Pauli feature
it allows to encode rectangular r × c images. Moreover, the and Zfeature map based quantum SVM kernel, as introduced in
number of qubits necessary to encode the image are further Havlicek et al [13]. In preparation of quantum kernel matrices
reduced. The QPIE representation is expressed as: for training and testing, feature map is first applied to each
22n −1
pair of training datapoints, then to testing datapoints. In later
X stage, train and test quantum kernel matrices are utilised for
|Ii = ci |ii, n = [log2 (rc)] (6)
i=0
classification support vector machine.
C. Classical Edge and Corner Detection realized by calculating each pixel’s gradient. It is based
1) Edge Detection: Edge detection is an image processing on the local auto-correlation function of a signal which
technique for finding the boundaries of an object in the given measures the local changes of the signal with patches
image. The edges are the part of the image that represents shifted by a small amount in different directions.
the boundary or the shape of the object in the image. the 3) Förstner corner detector [20]: This method solves for
general method of edge detection is to study the changes of the point closest to all the tangent lines of the corner in
a single image pixel in a gray area, use the variation of the a given window and is a least-square solution in order
edge neighboring first-order or second-order to detect the edge. to achieve an approximate solution. The algorithm relies
There are various methods in edge detection, and following are on the fact that for an ideal corner, tangent lines cross
some of the most commonly used methods: at a single point.
4) SUSAN (Smallest Uni-value Segment Assimilating
1) Prewitt edge detection [14]: This method is commonly
Nucleus) corner detector [21]: This method is based on
used to detect edges based applying a horizontal and ver-
brightness comparison and realized by a circular mask.
tical filter in sequence. This gradient based edge detector
If the brightness of each pixel within a mask is compared
is estimated in the 3x3 neighborhood for eight directions.
with the brightness of that mask’s nucleus, then an area
All the eight convolution masks are calculated.
of the mask can be defined which has the same (or
2) Sobel edge detection [15]: Sobel detection finds edges
similar) brightness as the nucleus. The SUSAN area will
using the Sobel approximation to the derivative. It
reach a minimum while the nucleus lies on a corner
precedes the edges at those points where the gradient
point.
is highest. It is one of the most commonly used edge
5) Fuzzy System [22]: The measure of “cornerness” for
detectors and helps reduce noise and provides differen-
each pixel in the image is computed by fuzzy rules
tiating, giving edge response simultaneously.
(represented as templates) which are applied to a set of
3) Laplacian edge detection [16]: This method uses only
pixels belonging to a rectangular window. The possible
one filter (also called a kernel). In a single pass, Lapla-
uncertainty contained in the window-neighborhood is
cian edge detection performs second-order derivatives
handled by using an appropriate rule base (template set).
and hence are sensitive to noise. The Laplacian is
The fuzzy system is simple to implement and still fast
generally used to found whether a pixel is on the dark
in computation when compared to some existing fuzzy
or light side of an edge.
methods. Also, it can be easily extended to detect other
4) Canny edge detection [17]: This is the most commonly
features.
used highly effective and complex compared to many
other methods. It is a better method that without disturb- D. Quantum Methods in Edge Detection
ing the features of the edges in the image afterwards and In the known classical edge extraction algorithms, for a typ-
it apply the tendency to find edges and serious value of ical image of 2n x 2n processing work with the computational
threshold. The algorithmic steps are as follows: complexity less than O(22n ), it is impossible to complete such
a) Convolve an image f (r, c) with a Gaussian func- a task [23]. By contrast, quantum image processing utilizes
tion to get smooth image f (r, c).f (r, c) = f (r, c)∗ the superposition and entanglement characteristics of quantum
G(r, c, 6). mechanics to simultaneously calculate all the pixels of the im-
b) Apply first difference gradient operator to compute age, thus realizing the acceleration of the algorithm. Basically,
edge strength, then edge magnitude and direction the quantum edge detection methods are the combination of
are obtained as before. QIR methods and classical edge methods.
c) Apply non-maximal or critical suppression to the 1) QSobel and other quantum image model based
gradient magnitude. methods: Early in 2014, Zhang et al. reported a novel
d) Apply threshold to the non-maximal suppression quantum image edge extraction algorithm called QSobel
image. [24]. It is designed based on the FRQI and the famous
2) Corner Detection: Corner detection works on the prin- edge extraction algorithm Sobel. Fan et al. proposed
ciple that if a small window is placed over an image, and that an image edge detection method based on Laplacian
window is placed on a corner. If that window is moved in any operator and Sobel in 2019 [25]. In 2020, Xu et al.
direction, then there will be a large change in intensity. proposed a quantum image processing algorithm us-
1) The Moravec corner detection [18]: This is one of the ing edge extraction based on Kirsch operator [23]. A
earliest corner detection algorithms and defines a corner quantum edge detection algorithm based on improved
to be a point with low self-similarity. The algorithm tests eight-direction Sobel operator was proposed by Ma et
each pixel in the image to see if a corner is present, by al. in 2020 [26]. The flow chart is shown in the Figure
considering how similar a patch entered on the pixel is 4, which is divided into 4 stages more specifically.
to nearby, largely overlapping patches. The classical digital image is quantized into quantum
2) Harris corner detector [19]: This method was devel- image NEQR model first. Then, the X-Shift and Y-Shift
oped to identify the internal corners of an image and it is transformations are used to obtain the shifted image set.
 Following that, very pixel’s gradient is further calculated
according to Sobel mask using the shifted image set    
simultaneously. Finally, the edge of the original image c0 c0 + c 1
is extracted through the threshold operation UT .

 c1 


 c0 − c1 

 c2   c2 + c 3 

c3
 1  c2 − c3

(I2n−1 ⊗ H) ·  → √  .
   
 ..  2  .. 

 . 


 . 

 cN−2   cN−2 + cN−1 
cN−1 cN−2 − cN−1
(10)
It is obvious that we now have access to the gradient
between the pixel intensities of neighboring pixels in
the form of ci − ci+1 . The logic here is: When two
pixels belong to the same region, their intensity values
are identical and the difference vanishes, otherwise
their difference is non-vanishing, which indicates a
region boundary. This process results in the detection
Fig. 4. Basic workflow of quantum image edge extraction, image from Fan of horizontal boundaries between the even-pixels-pairs
et al.’s work [25]
0/1, 2/3, and so on. For detection of horizontal bound-
aries between odd-pixel-pairs 1/2, 3/4, etc., we can per-
form an amplitude permutation on the quantum register
to convert the amplitude vector (c0 , c1 , c2 , . . . , cN −1 )T
2) Quantum Hadamard Edge Detection (QHED): An-
to (c1 , c1 , c2 , . . . , cN −1 , c0 )T , and then applying the H-
other kind of quantum edge detection algorithm cal-
gate and measuring the quantum register conditioned on
culated the difference between adjacent pixels in the
LSB being |1i.
quantum image via Hadamard transform, this method
is called Quantum Hadamard Edge Detection (QHED),
which is reported by Yao et al. in 2017 [27]. This QHED E. Image Classification
algorithm encode the pixel values of the image in the 1) Support Vector Machine(SVM) and Quantum SVM:
probability amplitudes and the pixel positions in the SVM classification is essentially a binary (two-class)
computational basis states. In a N-pixel image, the pixels classification technique and kernel methods can achieve
of the image are numbered in the form of |b1 b2 ...bn−1 0> more complexity in classical SVM. Similar to support
,where bi = 0 or 1. The positions of any pair of vector machines, the Quantum SVM algorithm applies
neighboring pixels in a picture column are given by the to classification problems that require a feature-map
binary sequences |b1 b2 ...bn−1 0 > and |b1 b2 ...bn−1 1 >, implicitly specified by a kernel. Quantum SVM using
and the corresponding pixel intensity values are stored minimization quantum subroutine GroverOptim and pro-
as the cb1 b2 ...bn−1 0 and cb1 b2 ...bn−1 1 . By applying the vides convergence to a global optimum for non-convex
Hadamard transform to the least significant bit on an cost function. Quantum SVM and kernels can exploit the
arbitrary size quantum register. The total operation is higher dimensional space efficiently and can generate
then, feature maps and subsequent decision boundaries that
are difficult for classical kernel functions to match.
As it mentioned in Park et al.’s work [28], Quantum
SVM can potentially provide better performance if the
underlying boundary is a complex one not captured well
  by traditional classical kernels.
1 1 0 0 0 0
2) K-means and Quantum K-means: K-means algo-
 1 −1 0 0 ... 0 0
  rithm belongs to the family of unsupervised machine
 0 0 1 1 ··· 0 0
1   learning algorithms that groups together n observations
I2n−1 ⊗H= √ 
 0 0 1 −1 0 0
 into K clusters making sure intra cluster variance is
2 .. ..  ..

 . . 
 . minimized. The most resource consuming operation for
 0 0 0 0 ... 1 1  k-means algorithm is calculation of a distance between
0 0 0 0 1 −1 vectors. The quantum version of the K-means algorithm
(9) provides an exponential speed-up for very high dimen-
where I2n−1 is the 2n−1 X 2n−1 unit matrix. sional input vectors by the fact that only log N qubits
Applying this unitary to a quantum register containing are required to load N-dimensional input vectors using
pixel values encoded using the QPIE representation, as the amplitude encoding [29].
 3) Convolutional Neural Network(CNN) and Quantum which 10000 images are used as test dataset. This dataset is
CNN: CNN has shown very high performance in com- divided in 6 batches, 5 batches are used for training the model
puter vision with the advantage of making good use of and 1 batch is used for model evaluation as a test dataset.
the correlation information of data. However, CNN is
challenging to learn efficiently if the given dimension D. Synthetic Datasets
of data or model becomes too large. Cong extends the Current quantum devices (NISQ) have a limited number
convolution layer and the pooling layer,the main features of qubits, thereby, limiting the exploration of images with
of CNN, to quantum systems as a model of Quantum large dimensions, to create synthetic datasets for evaluation
CNN [30]. The convolution circuit will find the hidden of the quantum algorithms. Furthermore, we also plan to
state by applying multiple qubit gates between adjacent apply dimensionality reduction techniques such as Principal
qubits and the pooling circuit will reduce the size of Component Analysis (PCA) on publicly available datasets
the quantum system by observing the fraction of qubits before passing them through the quantum devices.
or applying 2-qubit gates such as CNOT gates. After
III. R ESULTS
repeating the convolution circuit and pooling circuit until
the size of the system is sufficiently small, the fully A. Classical Methods
connected circuit can predict the classification result. 1) Edge detection with MNIST dataset:
The construction of the quantum convolutional layer can • Stage 1: Normalization and Smoothing the image: An
be divided into four steps: (1) The encoding process example of this stage is shown in Figure 6.
stores the pixel data corresponding to the filter size in • Stage 2: Getting edge related information using gaussian,
qubits; (2) The learnable quantum circuits apply the laplace, sobel filter as demonstrated in Figure 7.
filters that can find the hidden state from the input state; • Stage 3: Harris corner detector to detect edges in the
(3) The decoding process gets new classical data by digits as demonstrated in Figure 9.
measurement; (4) Repeat steps (1) to (3) to complete • Stage 4: Conversion to bag-of-words representation.
the new feature map. [31] • Stage 5: Histogram representation.
• Stage 6: Classification using SVM.
II. DATASETS
A. MNIST Dataset
This dataset [32] is an extensive set of 60000 handwritten
digits, every image has a dimension of 28x28 pixels. Along
with training data it also provides 10000 additional test dataset
consisting of equally distributed digits from 0 − 9.
B. Fashion MNIST Dataset
This dataset [33] consists of 60000 clothing images, 50000
training data and 10000 test dataset, with label from 10
different classes, every image has a dimension of 28x28 pixels.
Here, for quantum binary classification, small dataset is used
for training and subsequent testing of 3 equally distributed
clothing data class labelled as,
Fig. 6. stage-1 : Normalization and Smoothing the image

2) Deep Learning Model for MNIST dataset: We use the

architecture with Resnet [35] blocks, Batch normalization and
ReLU activation for classification of MNIST dataset. The
architecture provides an accuracy of 99.6% in only 12 epochs.
We used the Learning Rate Finder provided by FastAI library
[36] and found that the best learning rate for our architecture
is 0.05. We used Stochastic Gradient Descent (SGD) as the
optimizer and the loss function as Cross Entropy Loss as
MNIST is a multi-class classification problem.
Fig. 5. Fashion MNIST data with label 0, 2, 3.
3) Deep Learning Model for CIFAR-10 dataset: Here we
have used SOFTMAX activation function, with 10 units in
C. CIFAR-10 Dataset output layers. In compiling of model, we took account for loss
This dataset (Learning Multiple Layers of Features from function using Sparse Categorical-Cross Entropy as classes are
Tiny Images) [34] consists of 60000 images distributed equally totally distinctive. We have used Adam Optimizer to adapt
across 10 different classes, each image is 32x32 pixels out of learning rate, as it updates weights after every iteration. After
Fig. 7. stage-2 : Getting edge related information using gaussian, laplace,
sobel filter.

Fig. 10. Deep learning model summary (CNN)

Fig. 11. Deep learning model results (CNN)

analysis and normalization. For multi-class classification we

have used three quantum support vector machines based binary
Fig. 8. stage-3 : Harris corner detector to detect edges in the digits.
classifiers with One-vs-Rest approach classifying a particular
label e.g., label 0 as positive (1) and rest as negative (0).
In second stage, classical data was encoded to quantum state
using different quantum feature Map, followed by preparing
quantum kernel matrices utilising quantum state-vector simu-
lator for training and validation dataset. Finally, accuracy of
quantum kernel matrices was estimated on validation and test
dataset with prediction probability of particular label for test
data.
Finally, the result was obtained using the different quantum
feature map based quantum kernels in correctly identifying and
Fig. 9. Confusion Matrix and Prediction accuracy on Test dataset of MNIST
predicting label in the test data based on the highest probability
corresponding to a particular label. The Quantum binary image
running 15 epochs, final results obtained with accuracy of classification was performed based on different feature map i.e.
77.5%. ZZFeature, Pauli and ZFeature map results were compared for
the obtained accuracy of classifying different image labels.
B. Quantum Methods Quantum kernel based on ZZfeature map involves entan-
1) Quantum Image binary classification of MNIST dataset: gling blocks in different configuration i.e linear, circular and
Here, IBM Quantum lab platform is used for binary classifica- full entanglement for encoding classical data. This comparison
tion of fashion MNIST image data. [37] In the method, at first shows prediction accuracy within ZZfeature map with different
classical data was split into sample, validation and test data and entanglement configuration of test image data labels, which
labels, followed by preprocessing classical dataset with fea- helped in realization of achieving suitable configuration for
ture dimension 5 using standardization, principal component 3-class image classification with high accuracy.
 Fig. 15. Prediction based on different entanglement configuration of ZZFea-
ture map

Fig. 12. Prediction probability of Image label based on different feature map
based edge detection. The time complexity of classical op-
eration is O(2n ) however the quantum system can perform
the gradient calculation between pixels much faster using
hadamard operation O(mn.log(mn)). Hence leveraging the
faster operating speed this method performed edge detection
Fig. 13. Quantum Binary Image classification: Test result Label 0, 2 ,3
using quantum circuits and performed classification using
Logistic Regression was used to get the accuracy of the model.
For image representation, QPIE (Quantum Probability Image
Encoding) is used, where the image is using n = log2 N qbits
to represent an image where N is the number of pixels from
the image.
Results:

Fig. 16. QPIE image representation

The Algorithm performed two scan for gradient calcula-

Fig. 14. Different feature map based quantum kernel accuracy in prediction tions, in first scan only horizontal edges were detected and in
second scan the vertical edges were detected. This operation
is equivalent to SobelX and SobelY operation in classical
2) Quantum Image Classification using Edge detection : algorithms. The method, also handles large images by splitting
This method is split into two stages, the first stage includes them in smaller patches and applies edge detection separately.
finding edges of the image and in the second stage the edges This is majorly done due to limitation of qbits in actual
are then used for image classification. The edge detection with quantum computers, coupling between qbits and also to reduce
classical algorithms is much slower compared to hadamard the probability of noise impacting the operation. For MNIST
 Fig. 20. The flow of QHED with large images.
Fig. 17. Quantum Hadamard Edge detection using Simulator

Fig. 21. Accuracy of Method2 with sample MNIST dataset

Fig. 22. Accuracy of Method2 with sample CIFAR dataset

Fig. 18. Quantum Hadamard Edge detection horizontal scan 3) Impact of Noise: In the Noisy Intermediate-Scale Quan-
tum (NISQ) era, noise in quantum computers plays a crucial
role. We study the impact of various noise models for the
Quantum Edge Detection experiment on MNIST dataset.
• Depolarizing Noise: The depolarizing channel is defined
as the following:
I
E(ρ) = (1 − λ)ρ + λ Tr[ρ] (11)
2n
with 0 ≤ λ ≤ 4n / (4n − 1), where λ is the depolar-
izing error parameter and n refers to the number of
qubits. Depolarizing Noise creates spontaneous transi-
tions among eigenstates, inducing the sudden death of
Fig. 19. Quantum Hadamard Edge detection Vertical scan maximally entangled states [38]. We analyze the impact
of depolarizing noise on single qubit gates, two-qubit
gates and the combined effect on both types of gates.
dataset, the image is scaled to 32x32 pixels and then divided Figure 23 and 24 demonstrate the impact of depolarizing
into 4x4 pixel patches, the reason for choosing 32 is to provide noise with various noise probabilities. We notice that
support for CIFAR dataset and also increase the quality of noise on both the gates combined has a larger impact
edge detection because of up scaling. Coming to stage 2 of on the performance of the model.
the method, classification is done using a logistic regression • Pauli Noise: The amplitude and the phase of the qubit flip
model with mutli-class model including intercept term with based on the noise parameter provided to the Pauli Noise
Lasso regression as penalty. The Cifar dataset is more complex Model. We analyze the impact of bit flip noise, phase flip
compared to the MNIST dataset hence requires more time for noise and the combined effect of both the noises on the
edge detection, To perform training the scope was reduced by Quantum Edge Detection for MNIST dataset. Figure 25
using 100 samples to train from training dataset and the testing demonstrate the performance of the model with various
was performed using 10 samples of test dataset. which resulted Pauli Noise probabilities. We notice that noise probability
in 10% accuracy. The reduce in accuracy can be linked to the 0.001 has a large impact on the model’s performance,
incorrect splitting ratio of the image to small samples which decreasing the accuracy from 0.74 to 0.68.
was used to detect the edges and also due to smaller samples Figure 26 compares the edge detection results for model
of training. with no noise, depolarizing noise and pauli noise. We
 of the edge detector classifier with training data from
depolarizing noise model and pauli noise model. Pauli
Noise has a worse impact on the model’s accuracy.

Fig. 23. Impact of Depolarizing Noise with a Noise Probability of 0.001,

0.01

Fig. 26. Comparing the edge detection results for Noiseless model (top 3
images), Depolarizing Noise model with Noise Probability 0.5 for single
qubit and two-qubit gates (middle 3 images), Pauli Noise model with Noise
Probability 0.01 for both bit flip and phase flip (bottom 3 images)

Fig. 24. Impact of Depolarizing Noise with a Noise Probability of 0.1, 0.5

Fig. 27. Comparing the model performance with depolarizing noise and pauli
noise

IV. C ONCLUSION
In quantum image processing, we have used different im-
age encoding methods for edge detection and binary image
classification. When compared to classical process, quantum
provides advantage in runtime, space complexity i.e. number
of bits/qubits required for image encoding, but is not suffi-
cient in terms of depth and width of image. With Quantum
methods it is quite challenging to handle larger images as
Fig. 25. Impact of Pauli Noise with a Noise Probability of 0.001, 0.01 circuit design gets complex with increased number of qubits
required, leading to addition of noise with inaccurate edge
detection and classification results. In this project, we have
notice that after adding different noises, the digits are still used gate based quantum statevector and qasm simulator that
recognizable. Figure 27 demonstrates the performance are slower in performance which in-turn increases the training
time for larger datasets and requires splitting of data into [19] Chris Harris, Mike Stephens, et al. A combined corner and edge detector.
smaller sets and recombining to get higher accuracy for larger In Alvey vision conference, volume 15, pages 10–5244. Citeseer, 1988.
[20] Wolfgang Förstner and Eberhard Gülch. A fast operator for detection
image pixel and datasets. While running real quantum may and precise location of distinct points, corners and centres of circular
provides better performance but it is more exposed to the features. In Proc. ISPRS intercommission conference on fast processing
noise from the devices. As advancements with this projects, of photogrammetric data, pages 281–305. Interlaken, 1987.
[21] Stephen M Smith and J Michael Brady. Susan—a new approach to
our team is working to analyse impact of different noise low level image processing. International journal of computer vision,
models on the quantum image encoding circuit and its results 23(1):45–78, 1997.
when processing smaller pixel as well as large pixel images. [22] Minakshi Banerjee and Malay K Kundu. Handling of impreciseness in
gray level corner detection using fuzzy set theoretic approach. Applied
Our work in future will include implementing from quantum Soft Computing, 8(4):1680–1691, 2008.
processed results for image verification with original/real-time [23] Pengao Xu, Zhenxing He, Tianhui Qiu, and Hongyang Ma. Quantum
images to reach better accuracy. image processing algorithm using edge extraction based on kirsch
operator. Optics express, 28(9):12508–12517, 2020.
[24] Yi Zhang, Kai Lu, and YingHui Gao. Qsobel: A novel quantum image
R EFERENCES edge extraction algorithm. Science China Information Sciences, 58, 12
[1] Yue Ruan, Xiling Xue, and Yuanxia Shen. Quantum image processing: 2014.
Opportunities and challenges. Mathematical Problems in Engineering, [25] Ping Fan, Ri-Gui Zhou, Wen Wen Hu, and NaiHuan Jing. Quantum
2021, 2021. image edge extraction based on laplacian operator and zero-cross
[2] Fei Yan, Abdullah M Iliyasu, and Salvador E Venegas-Andraca. A method. Quantum Information Processing, 18(1):1–23, 2019.
survey of quantum image representations. Quantum Information Pro- [26] Yulin Ma, Hongyang Ma, and Pengcheng Chu. Demonstration of
cessing, 15(1):1–35, 2016. quantum image edge extration enhancement through improved sobel
[3] Salvador E Venegas-Andraca and Sougato Bose. Storing, processing, operator. IEEE Access, 8:210277–210285, 2020.
and retrieving an image using quantum mechanics. In Eric Donkor, [27] Xi-Wei Yao, Hengyan Wang, Zeyang Liao, Ming-Cheng Chen, Jian
Andrew R. Pirich, and Howard E. Brandt, editors, Quantum Information Pan, Jun Li, Kechao Zhang, Xingcheng Lin, Zhehui Wang, Zhihuang
and Computation, volume 5105, pages 137 – 147. International Society Luo, and et al. Quantum image processing and its application to edge
for Optics and Photonics, SPIE, 2003. detection: Theory and experiment. Physical Review X, 7(3), Sep 2017.
[4] Phuc Q Le, Fangyan Dong, and Kaoru Hirota. A flexible representation [28] Jae-Eun Park, Brian Quanz, Steve Wood, Heather Higgins, and Ray
of quantum images for polynomial preparation, image compression, and Harishankar. Practical application improvement to quantum svm: theory
processing operations. Quantum Information Processing, 10(1):63–84, to practice. arXiv preprint arXiv:2012.07725, 2020.
2011. [29] Dawid Kopczyk. Quantum machine learning for data scientists. arXiv
[5] Panchi Li, Hong Xiao, and Binxu Li. Quantum representation and preprint arXiv:1804.10068, 2018.
watermark strategy for color images based on the controlled rotation [30] Iris Cong, Soonwon Choi, and Mikhail D Lukin. Quantum convolutional
of qubits. Quantum Information Processing, 15(11):4415–4440, 2016. neural networks. Nature Physics, 15(12):1273–1278, 2019.
[6] Rabia Amin Khan. An improved flexible representation of quantum [31] Seunghyeok Oh, Jaeho Choi, and Joongheon Kim. A tutorial on quantum
images. Quantum Information Processing, 18(7):1–19, 2019. convolutional neural networks (qcnn). In 2020 International Conference
[7] Yi Zhang, Kai Lu, Yinghui Gao, and Mo Wang. Neqr: a novel en- on Information and Communication Technology Convergence (ICTC),
hanced quantum representation of digital images. Quantum information pages 236–239. IEEE, 2020.
processing, 12(8):2833–2860, 2013. [32] Yann LeCun and Corinna Cortes. MNIST handwritten digit database.
[8] RiGui Zhou, WenWen Hu, GaoFeng Luo, XingAo Liu, and Ping Fan. 2010.
Quantum realization of the nearest neighbor value interpolation method [33] Han Xiao, Kashif Rasul, and Roland Vollgraf. Fashion-mnist: a novel
for ineqr. Quantum Information Processing, 17(7):1–37, 2018. image dataset for benchmarking machine learning algorithms. CoRR,
[9] Hai-Sheng Li, Xiao Chen, Shuxiang Song, Zhixian Liao, and Jianying abs/1708.07747, 2017.
Fang. A block-based quantum image scrambling for gneqr. IEEE Access, [34] Alex Krizhevsky, Vinod Nair, and Geoffrey Hinton. Cifar-10 (canadian
7:138233–138243, 2019. institute for advanced research).
[10] Simona Caraiman and Vasile Manta. Image representation and process- [35] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep
ing using ternary quantum computing. In International Conference on residual learning for image recognition, 2015.
Adaptive and Natural Computing Algorithms, pages 366–375. Springer, [36] Jeremy Howard and Sylvain Gugger. Fastai: A layered api for deep
2013. learning. Information, 11(2):108, Feb 2020.
[11] Hai-Sheng Li, Qingxin Zhu, Ri-Gui Zhou, Lan Song, and Xing-jiang [37] Quantum edge detection - qhed algorithm on small and large images.
Yang. Multi-dimensional color image storage and retrieval for a normal Qiskit Documentation.
arbitrary quantum superposition state. Quantum Information Processing, [38] Ruchipas Bavontaweepanya. Effect of depolarizing noise on entangled
13(4):991–1011, 2014. photons. In Journal of Physics: Conference Series, volume 1144, page
[12] Xi-Wei Yao, Hengyan Wang, Zeyang Liao, Ming-Cheng Chen, Jian Pan, 012047. IOP Publishing, 2018.
Jun Li, Kechao Zhang, Xingcheng Lin, Zhehui Wang, Zhihuang Luo,
et al. Quantum image processing and its application to edge detection:
theory and experiment. Physical Review X, 7(3):031041, 2017.
[13] Vojtěch Havlı́ček, Antonio D. Córcoles, Kristan Temme, Aram W.
Harrow, Abhinav Kandala, ’Jerry M. Chow, and Jay M. Gambetta.
Supervised learning with quantum-enhanced feature spaces. Nature,
567(7747):209–212, Mar 2019.
[14] John Francis Canny. Finding edges and lines in images. Technical report,
MASSACHUSETTS INST OF TECH CAMBRIDGE ARTIFICIAL
INTELLIGENCE LAB, 1983.
[15] FG Irwin et al. An isotropic 3x3 image gradient operator. Presentation
at Stanford AI Project, 2014(02), 1968.
[16] Tamar Peli and David Malah. A study of edge detection algorithms.
Computer graphics and image processing, 20(1):1–21, 1982.
[17] John Canny. A computational approach to edge detection. IEEE
Transactions on pattern analysis and machine intelligence, (6):679–698,
1986.
[18] Hans Peter Moravec. Obstacle avoidance and navigation in the real
world by a seeing robot rover. PhD thesis, Stanford University, 1980.