Quantum mechanics-based signal and image

denoising
Sayantan Dutta1,2 , Adrian Basarab1 , Bertrand Georgeot2 , and Denis Kouamé1
1 IRIT UMR CNRS 5505, University of Toulouse, Université Paul Sabatier, CNRS, France
2 Laboratoirede Physique Théorique, IRSAMC, Université de Toulouse, CNRS, UPS, France

Abstract—Decomposition of digital signals and images into literature. One may note the seminal work in [6], or, more
arXiv:2004.01078v1 [eess.SP] 2 Apr 2020

other basis or dictionaries than time or space domains is a recently, the interest of quantum mechanics in image segmen-
very common approach in signal and image processing and tation [7] or in pulse-shaped signal analysis [8], [9]. Note that
analysis. Such a decomposition is commonly obtained using fixed
transforms (e.g., Fourier or wavelet) or dictionaries learned from a separate domain also exists on designing image processing
example databases or from the signal or image itself. In this work, algorithms adapted to quantum computers, but is of a different
we investigate in detail a new approach of constructing such a purpose [10], [11].
signal or image-dependent bases inspired by quantum mechanics More related to our work, we note that there was a recent
tools, i.e., by considering the signal or image as a potential in attempt to use quantum mechanics in the same context in
the discretized Schroedinger equation. To illustrate the potential
of the proposed decomposition, denoising results are reported in [12], [13]. Although there are similarities between the two
the case of Gaussian, Poisson and speckle noise and compared approaches, there are also some important differences. The
to state of the art algorithms based on wavelet shrinkage, total authors in [12], [13] start from a continuous mathematical
variation regularization or patch-wise sparse coding in learned representation of the signal, and the discretization only occurs
dictionaries. at the end of the process. The processing of a large image in
Index Terms—adaptive signal and image representation, de- these papers is done by decomposing it into lines and columns
noising, quantum mechanics. to get 1D signals, while the proposed work is applied block-
wise, which we believe is more efficient for image denoising.
Additionally, unlike [12], [13], our method fully takes into
I. I NTRODUCTION
account the quantum localization phenomenon, a subtle effect
In number of applications, processing or analyzing signals due to quantum interference which makes the distribution
and images require the use of other representations than of the eigenfunctions of the Schroedinger operator strongly
time or space. While the most famous transformation still dependent on noise, and has important effects on the denoising
remains the Fourier transform, other representations have been process. We also use the physics of the problem to identify
proposed to overcome the non-localization in time or space the optimal domain of applicability of such methods.
of the Fourier basis vectors. Among these time-frequency The proposed framework reposes on the discrete version of
representations, the most used are the short time Fourier and the Schroedinger equation for a quantum particle in a potential.
the wavelet transforms [1], [2]. Most often (see, e.g., image In our case, the potential is represented by the signal samples
compression, restoration, reconstruction, denoising or com- or the pixel values. The bases used to decompose the signal
pressed sensing), such transforms are associated to the concept or the image are directly computed from the signal and image
of sparsity, i.e., their ability to concentrate most of the signal or itself and correspond to the wave function representing the
image energy in a few coefficients. To reinforce the sparsity, stationary solutions of the Schroedinger equation. These wave
overcomplete dictionaries have also been explored over the functions have interesting properties of temporal or spatial
last decades, such as the wavelet frames or more recently localization and of frequency dependence on the value of the
patch-based or convolutional dictionaries learned from a set potential. In particular, they present higher frequencies for low
of training signals or images [3]. The latter has been shown potential values, thus allowing an original signal or image
to be of particular interest in image denoising [4]. decomposition.
In this paper, we investigate a novel signal and image repre- Within the proposed framework, the frequency and lo-
sentation, through a dedicated basis extracted from the signal calization properties of the basis can be controlled through
or image itself, using concepts from quantum mechanics. First several parameters, thus ensuring flexibility in applications
preliminary results were published in [5]. Compared to fixed such as denoising. A detailed description of the behavior of
basis such as Fourier, discrete cosinus, wavelets, curvelets, etc, the proposed transform and denoising method with respect to
or dictionary learning that generally needs a training database, the choice of these parameters is provided, allowing to gain
the proposed approach has the advantage of computing a insight about the practical consequences in signal and image
transform adapted to the signal or image of interest. processing of the quantum mechanical principles involved.
Several attempts of translating quantum principles in image Furthermore, the proposed transform embedded in a denoising
or signal processing applications have been proposed in the algorithm shows promising results in different noise scenarios
 2 2
(additive Gaussian, Poisson or speckle noise). Finally using H ∈ RN ×N defined asumming that the signal or the image
different signals and images, comparisons with several state x is the potential:
of the art methods are finally performed. 
~2
The remainder of the paper is organized as follows. Section  x(i, j) + 4 2m
 f or i = j,
~2

− 2m f or i = j ± 1,

II gives the details of the adaptive transform design and its H(i, j) = (2)
~2
application to denoising. Results and comparisons are provided 

 − 2m f or i = j ± N,
0 otherwise.

in Section III and concluding remarks are finally reported in
Section IV. As the boundary conditions correspond to zero padding of
the image, a few individual coefficients of the matrix H follow
II. T HE METHOD ~2
specific rules. Indeed, H(i, j) = x(i, j) + 2 2m for i = j and
~2
A. Adaptive transform from quantum mechanics i ∈ {1, N, N 2 −N +1, N 2 }, H(i, j) = x(i, j)+3 2m for i = j
The main idea of the proposed method is to describe a signal and i ∈ {2, 3, ..., N − 1, N 2 − N + 2, N 2 − N + 3, ..., N 2 − 1},
~2
or an image on a specific basis which is constructed through H(i, j) = x(i, j) + 3 2m for i = j and other than the previous
the resolution of a related problem of quantum mechanics: the set with i mod N ∈ {0, 1} and H(i, i + 1) = H(i + 1, i) =
probability of presence of a quantum particle in a potential 0 for any i multiple of N apart from N 2 . As an example,
related to the signal or image. In the following, we explain for an image of size 4 × 4 (i.e. N = 4), the corresponding
how this problem is formulated and how it is built from the Hamiltonian matrix will be of size 16 × 16 (explicitly shown
signal or image. in Appendix A).
Our formalism is based on the resolution of the The set of eigenvectors gives a basis of the Hilbert space,
Schroedinger equation of non-relativistic quantum mechan- with each eigenvector associated to an energy E, which is the
ics. This equation determines the wave function ψ(y) which corresponding eigenvalue of the Hamiltonian operator. The N 2
2
belongs to the Hilbert space of L2 -integrable functions, y eigenvectors, denoted by ψ i ∈ RN ×1 are the main tool for the
being e.g. a spatial coordinate. The function |ψ(y)|2 gives proposed adaptive transform in this work. Indeed, our method
the probability of presence of the particle, which implies consists in projecting the signal or image on this particular
that |ψ(y)|2 dy = 1. For particles in a potential V (y), the
R
basis and use the energy associated to each eigenfunction as
stationary Schroedinger equation reads a parameter on which we perform the thresholding of these
coefficients.
~2 2
− ∇ ψ = −V (y)ψ + Eψ, (1)
2m B. Properties of the adaptive transform
with m the mass of the particle and ~ the Planck constant The basis vectors are ranked from the lowest to the highest
that are parameters of the problem. In operator notation it corresponding eigenvalue, associated with the energy E. In
~2
corresponds to Hψ = Eψ with H = − 2m ∇2 + V the the same way as the Fourier or wavelet basis, these basis
Hamiltonian operator. The energy E of the particle in (1) vectors ψ i are oscillating functions,
labels the solutions of the problem. Solutions of this stationary p but with a local frequency
depending on the local value of 2m(E − V )/~. This local
Schroedinger equation in a bounded domain correspond to a frequency thus depends on the difference between the energy
discrete set of energy levels, from a minimal energy to infinity. E and potential V , such that for a given energy, the higher
Solutions of (1) form a basis of the Hilbert space to which frequency oscillations are associated with the lower values of
the wavefunctions belong. This space is infinite-dimensional the potential. In this way, the basis of eigenvectors of (1) natu-
for continuous values of the position y in (1). However, rally describes with different frequencies the different parts of
we are interested in signal or image processing applications, the signal or image, in contrast to e.g the Fourier or wavelet
where the space is discretized in a finite number of points. bases. The precise relation between the local frequency of the
Specifically, we assume that the potentiel V is represented by eigenvectors and the value of the signal or image pixel is
the value of signal sample or image pixel x. In the case of governed by the parameter ~2 /2m. In the physical problem
a discretized problem, the operators become finite matrices of quantum mechanics, this quantity is linked to Planck’s
and the resolution of (1) is equivalent to diagonalizing the constant and the particle mass, but in our framework it is a free
Hamiltonian matrix. In (1), the Laplacian operator should parameter. It should be chosen with care, as extreme values
be replaced by its discrete version, following the standard are clearly inadequate. Indeed, as the problem is discretized
numerical definitions of the gradient operator, developed in there is a maximal frequency in the problem, linked to the
the following for a 2D image x ∈ RN ×N : inverse of the discretization step. If ~2 /2m is very small, the
p
(∂h x)(i, j) = x(i + 1, j) − x(i, j) if i < N local frequencies 2m(E − V )/~ become very high even for
low values of the energy, the maximal energy becomes very
(∂v x)(i, j) = x(i, j + 1) − x(i, j) if j < N
low, and the basis does not explore properly high values of the
where ∂h and ∂v are associated to the horizontal and vertical signal or pixels of the image. On the other side for very large
gradients. The boundary conditions correspond simply to a values of ~2 /2m, the local frequencies become smaller and
zero padding of the image. smaller at fixed energy, the maximal energy becomes larger
The resolution of (1) is thus equivalent to finding eigen- and larger, and eventually when ~2 /2m tends to infinity most
vectors and eigenvalues of the discretized Hamiltonian matrix vectors of the adaptive basis are so high above the signal or
Fig. 1. Pictorial diagram of the proposed denoising framework.

image pixel that they do not discriminate between low and image textures without generating artifacts. The most common
high values, becoming closer and closer to the standard Fourier denoising strategies are based on three primary steps. To
basis vectors. Therefore it is crucial to tune the free parameter distinguish the useful information and the noise the noisy
~2 /2m in the right way. signal or image is projected onto a dictionary. This is then
A more subtle property of the basis consists in its local- accompanied by a hard or soft thresholding process in the
ization properties. Indeed, it is known in quantum mechanics transformed space. Finally, the revised coefficients are back
that a disordered potential localizes the wavefunctions in one projected to the time or space domain, so that the denoised
and two dimensions. Due to destructive interference, the dif- signal or image could be retrieved. We will apply the same
ferent wave functions are exponentially localized at different procedure using the adaptative basis defined by the eigenvec-
positions of the potential, an effect known as Anderson local- tors ψ i obtained by solving the Schroedinger equation (1).
ization, which earned the Nobel prize in 1977 to its discoverer The basic assumptions is that the noise is more present in
[14]. If the signal or image are not smooth, which certainly high frequency components of the signal or image, correspond-
arises in the case of a noisy signal or image, we expect the ing to eigenvectors associated with large energy eigenvalues.
vectors of the basis to be localized, with a localization length The thresholding will therefore be performed in energy, leav-
which will be smaller and smaller for increasing noise power. ing out the components of the signal or image on high energy
The level of localization is measured by computing the eigenvectors. The fact that our basis has frequencies which
inverse participation ratio (IPR) of the wave functions, math- vary depending on the position should be an asset, especially
ematically defined as: for signal or image dependent noise (e.g. Poisson noise). In
PN 2 the following, we will show that it is indeed the case in some
|ψ i |2 examples of signals and images with various types of noise.
IP R = Pi=1 N 2 (3)
4 The denoising process unfolds as follows; for a noisy signal
i=1 |ψ i |
or image denoted by x, the denoised signal or image is rebuilt
where N 2 is the dimension of the Hilbert space. For a through:
vector uniformly spread over P indices and zero elsewhere, N2

this quantity is exactly P . For an exponentially localized x̂ =

X
αi ψ i τi , (4)
vector such as the wavefunctions in a disordered potential, i=1
it is proportional to the localization length. This phenomenon with
should be taken into account when producing the adaptative
basis. 
 1 f or i ≤ s,
i−s
τi = 1− ρ f or i > s and f or 1 − i−sρ > 0, (5)
C. Application to the denoising problem 
0 otherwise.
The significant difficulties for signal or image denoising where αi are the coefficients representing the signal or image
are to sharpen the edges without blurring and preserve the x in the proposed adaptive basis. s and ρ are two hyperparam-
eters, used to define the thresholding function for the proposed • Use algorithm 1 for each sub-block.
denoising algorithm. • Reconstruct the denoised signal or image by integrating
In order to use this procedure, we will need to specify each denoised sub-block.
which values of the parameter ~2 /2m should be selected.
As we will see, there is a relatively large range of values III. R ESULTS
where the algorithm is efficient, meaning that it can be set to This section regroups results showing the interest of the
a specific value independent of the signal or image on which proposed approach in signal and image denoising and analyze
the algorithm is used. the optimal choice of parameters. Subsection III-A illustrates
The localization property of the basis vectors should also be the localization property of the wave functions in the presence
taken into account. A given noisy signal or image introduced of noise, and consequently the interest of pre-smoothing with
in (1) will give rise to localized wave functions with a parameter σ embedded in the denoising algorithm. Subsection
localization length which will depend on the noise level. To III-B elaborates the dependence of the proposed denoising
modify this characteristic of the basis, and to construct the method on the choice of the hyperparameters ~2 /2m and σ.
Hamiltonian matrix, we use a smoothed adaptation of the noisy Subsection III-C compares the denoising results obtained with
signal or image, computed by a simple convolution with a the proposed approach to several state of the art methods.
Gaussian kernel whose standard deviation is denoted by σ. In Finally, the section ends with an example of real medical
our framework, this standard deviation σ is an additional free application in subsection III-D, showing the ability of the
parameter. If σ is chosen too large, then the noisy signal or proposed method to denoise real world (dental) cone beam
image becomes so smooth that many characteristics needed computed tomography (CBCT) images.
for the adaptative basis will be lost. On the opposite, if σ is
too small the basis vector will remain strongly localized. To 20

balance both sides we need to tune the parameter σ to get the 15 10

best achievable outcome. 20

10
Amplitude

30
5
D. Algorithm description 40
0
Denoising a signal or an image using the proposed method 50
-5
requires the computation of eigenvalues and eigenvectors of 60

the discretized Hamiltonian matrix (2) for appropriate values 0 100 200 300 400 500 10 20 30 40 50 60

of the parameters ~2 /2m and σ, project the signal or image (a) (b)
on this basis, threshold the coefficients by an appropriate
Fig. 2. (a) Synthetic signal and (b) cropped version of Lena used to illustrate
threshold in energy, and reconstruct from this a denoised signal the localization property of the wave functions.
or image. These steps are summarized in Algorithm 1 and Fig.
1.

Algorithm 1: Denoising algorithm using the proposed

100
adaptive transform.
~2 90
Input: x, 2m , s, ρ, σ
80
1 Compute a smooth version of x by Gaussian filtering
Construct the Hamiltonian matrix H based on the 70
IPR

2
smoothed version of x using (2) 60
3 Calculate the eigenvectors ψ i of H
4 Compute the coefficients αi by projecting x onto the 50
basis formed by ψ i
40
5 Threshold the coefficients αi and recover the denoised
signal or image following (5) and (4) 30

Output: x̂ 0 5 10 15 20
SNR (dB)

For very large signals and images, where the size of Fig. 3. IPR corresponding to the wave function calculated from the signal in
Fig. 2(a) degraded by an additive Gaussian noise for several SNR. The size
the matrix (2) becomes too large for practical simulations, of the signal was 512.
we implement a modified version of the algorithm where
the matrix (2) is diagonalized for subparts of the signal or
image independently, and then a complete signal or image is A. Localization of wave functions and interest of Gaussian
reconstructed: smoothing
• The noisy signal or image is divided into sub-blocks of The localization property of the wave functions and the
equal size. impact of Gaussian smoothing are analyzed through two
examples, i.e., a synthetic signal and a cropped version of Fig. 4(b), Fig. 5(b-c) and Fig. 6(b-c) show examples of wave
the standard Lena image shown in Fig. 2. functions calculated from a noisy signal and image, where
Fig. 3 shows the average IPR of solutions of (1) obtained the same noisy image corrupted by additive Gaussian noise is
from the signal in Fig. 2(a) degraded by an additive Gaussian considered in Fig. 5(a) and Fig. 6(a). The Fig. 5 and Fig.
noise with different signal to noise ratios (SNR). The results 6 in fact show two different wave functions for the same
displayed clearly highlight the increasing localization of the example. From these examples, one can observe that the wave
wave functions when the SNR decreases. This is consistent functions are completely localized in a specific location and
with the theory. Indeed, as explained in the previous section, present a fast decrease due to the destructive interference. On
due to the phenomenon of Anderson localization the wave the contrary, in the case where the same wave functions are
functions ψi are exponentially localized in the presence of calculated from low-pass filtered versions of the noisy signal
noise. In addition, they become more and more localized and image (i.e. a smoothed version of the potential), they are
when noise increases. This property can be measured by the shown to delocalize and spread over the whole available space
IPR defined in (3), which measures the number of significant using high frequencies for regions with low potential and low
components of a vector. frequencies for high potential value regions, as illustrated in

0.2 0.1

0.1 0.05
Amplitude

Amplitude
0 0

-0.1 -0.05

-0.2 -0.1

0 100 200 300 400 500 0 100 200 300 400 500

(a) (b) (c) (d)

Fig. 4. (a) Signal in Fig. 2(a) contaminated by additive Gaussian noise corresponding to a SNR of 15 dB, (b) localized wave function number 68 calculated
from the noisy signal with energy level illustrated by the dashed line in (a), (c) blurred version of the noisy signal in (a) obtained by Gaussian low-pass filter
corresponding to σ 2 = 10, (d) delocalized wave function number 68 calculated from the low-pass filtered signal with the same energy level illustrated by
the dashed line in (c).

10

20

30

40

50

60
10 20 30 40 50 60

(a) (b) (c)

10

20

30

40

50

60
10 20 30 40 50 60

(d) (e) (f)

Fig. 5. (a) Cropped Lena image in Fig. 2(b) contaminated by additive Gaussian noise corresponding to a SNR of 15 dB, (b,c) localized wave function number
25 calculated from the noisy lena image (a), (d) blurred version of the noisy lena image in (a) obtained by Gaussian low-pass filter corresponding to σ 2 = 6,
(e,f) the same wave function but delocalized due to the low pass Gaussian filter applied to the noisy image.
 10

20

30

40

50

60
10 20 30 40 50 60

(a) (b) (c)

10

20

30

40

50

60
10 20 30 40 50 60

(d) (e) (f)

Fig. 6. (a) Cropped Lena image in Fig. 2(b) contaminated by additive Gaussian noise corresponding to a SNR of 15 dB (which corresponds to the same
noisy Lena image used previously in Fig. 5(a)), (b,c) localized wave function number 195 calculated from the noisy lena image (a), (d) blurred version of
the noisy lena image in (a) obtained by Gaussian low-pass filter corresponding to σ 2 = 6, (e,f) the same wave function but delocalized due to the low pass
Gaussian filter applied to the noisy image.

Fig. 4(d), Fig. 5(e-f) and Fig. 6(e-f).

10 10

20 20
The localization of the wave functions in the presence of
30 30
noise has an important impact on the proposed signal or
40 40
image representation and furthermore on the efficiency of the
50 50
denoising process. To illustrate this claim, Fig. 7 shows a
60 60
denoising result with and without the use of the low pass
10 20 30 40 50 60 10 20 30 40 50 60
Gaussian filter prior to the computation of the wave functions.
(a) (b) In this example, the cropped version of Lena in Fig. 7(a) was
degraded by a Poisson noise resulting into a SNR of 15 dB.
The denoised images in Fig. 7(c,d) were obtained using the
10 10
algorithm detailed in Algo. 1. However, while the result in Fig.
20 20
7(c) exploits the image decomposition through localized wave
30 30 functions computed directly from the noisy image, the result in
40 40 Fig. 7(d) was obtained by filtering the noisy image by a low
50 50 pass Gaussian filter before using (1), in order to delocalize
60 60 the wave functions. The interest of such delocalization can be
10 20 30 40 50 60 10 20 30 40 50 60 visually appreciated in this example and allows a peak SNR
(c) (d) (PSNR) gain of more than 3 dB. In the following, we will
always use a pre-smoothed signal or image in (1), and the
Fig. 7. (a) Cropped version of clean Lena, (b) cropped version of noisy Lena parameter σ of the smoothing will be studied as one of the
contaminated by Poisson noise corresponding to a SNR of 15 dB, (c) denoised
result with PSNR = 25.37 dB of the image (b), (d) denoised result with PSNR parameters of the algorithm. To avoid confusion, let us insist
= 28.81 dB after smoothing the noisy image (b) by a low pass Gaussian filter on the fact that although the smoothing of the noisy signal
corresponding to σ 2 = 4. The hyperparameter ~2 /2m = 0.6 and the other or image before using (1) may look like a denoising process,
hyperparameters s and ρ have been manually tuned up to their best possible
values for each set of experiment. it is only a tool to construct the appropriate adaptative basis,
which will be used in a second step to denoise the original
noisy signal or image.
 B. Influence of hyperparameters on the efficiency of the algo-
0.2 rithm
Amplitude

0 In this subsection, we provide a detailed discussion about

the influence of the hyperparameters on the proposed adapative
-0.2
0 100 200 300 400 500 bases. As mentionned above, the parameter ~2 /2m specifies
how the local frequencies of the vectors of the basis vary with
Amplitude

0.1
the amplitude of the signal or image pixel value.
0
-0.1
0 100 200 300 400 500
20
0.1
Amplitude

-0.1
15
0 100 200 300 400 500

SNR
(a) 10
σ2 = 0
2
σ =4
0.1 5
σ2 = 8
Amplitude

0 σ2 = 40
-0.1 0
0 100 200 300 400 500 0 100 200 300 400 500
0.1 s
Amplitude

0
Fig. 9. Influence of the hyperparameter s on the reconstruction process of
-0.1 the denoised result of the 1D signal Fig. 2(a) corrupted by additive Gaussian
noise corresponding to a SNR of 15 dB, for ρ = 1, ~2 /2m = 0.4 and four
0 100 200 300 400 500
different values of σ 2 (0, 4, 8 and 40).
0.1
Amplitude

0 To illustrate this relationship, the effect of ~2 /2m on local

frequencies is shown in Fig. 8 for three distinct values of this
-0.1
parameter. For each case, three wave functions (number 25,
0 100 200 300 400 500
70 and 100) computed from the synthetic signal in Fig. 2(a)
are displayed. For low values of ~2 /2m (i.e., 0.08 for the
(b)
results in Fig. 8(a)), one may remark that the wave functions
are oscillating at very high frequencies, even for higher values
0.1 of the potential (i.e., of the signal). The presence of a maximal
Amplitude

oscillation period due to the discretization of the signal implies

0
that in this limit the high values of the signal are not taken
-0.1 properly into account. For very high values of ~2 /2m (15 for
0 100 200 300 400 500 the results in Fig. 8(c)), most of the wave functions are at an
0.1 energy well above the potential values, and they discriminate
Amplitude

less and less between the regions with different potential

0
height. In this limit, wave functions behave very similarly to
-0.1 cosine functions with increasing frequencies, thus reducing the
0 100 200 300 400 500 interest of the proposed bases that becomes very similar to
0.1 the Fourier transform. At intermediate values of ~2 /2m (1 for
Amplitude

0
the results in Fig. 8(b)), wave functions explore the different
regions but with clearly different oscillation frequencies, i.e.
-0.1 wave vectors have significantly larger frequencies or short
0 100 200 300 400 500
wavelengths for the low potential valued regions as opposed
to high potential regions.
(c)
The second hyperparameter studied in this section that
Fig. 8. Wave function number 25, 70 and 100 calculated from the signal Fig. has a strong impact on the proposed denoising algorithm is
2(a) are shown from top to bottom with the parameter (a) ~2 /2m = 0.08, the cut-off frequency of the Gaussian low pass filter used
(b) ~2 /2m = 1 and (c) ~2 /2m = 15.
to smooth the noisy signal or images before computing the
wave functions through (1), as explained in the preceding
 (a) (b)

(c) (d)

(e) (f)

Fig. 10. Influence of the hyperparameters ~2 /2m and σ on proposed decomposition performed on the 1D system Fig. 2(a) in presence of (a,b) Poisson
noise, (c,d) Gaussian noise and (e,f) speckle noise corresponding to a SNR of 15 dB respectively. The hyperparameters s and ρ have been manually tuned
up to their best possible values for each set of experiment.

subsection. In particular, this cut off frequency is fixed through At last, in order to denoise the signal or image one has to
the choice of the standard deviation σ of the Gaussian filter. threshold the coefficients of the signal or image on the adapta-
 20

15

10
Amplitude

-5

0 100 200 300 400 500

(a) (b) (c)

20

15
Amplitude

10

0 100 200 300 400 500

(d) (e) (f)

40

30
Amplitude

20

10

0 100 200 300 400 500

(g) (h) (i)

Fig. 11. (a) Sample A, (b,c) influence of the hyperparameters ~2 /2m and σ on proposed method performed on the sample A corrupted with Poisson noise
corresponding to a SNR of 15 dB, (d) sample B, (e,f) influence of the hyperparameters ~2 /2m and σ on proposed method performed on the sample B
corrupted with Poisson noise corresponding to a SNR of 15 dB, (g) sample C, (h,i) influence of the hyperparameters ~2 /2m and σ on proposed method
performed on the sample C corrupted with Poisson noise corresponding to a SNR of 15 dB. The hyperparameters s and ρ have been manually tuned up to
their best possible values for each set of experiment.

tive basis; this process uses two thresholding hyperparameters the smoothing process before calculating the wave functions
s and ρ defined in (5), which define respectively the threshold through (1). For nonzero values of σ 2 , there is a relatively
value and the abruptness of the cut off. In particular, the small range of optimal s values, where the SNR is much better
parameter s corresponds to the threshold in energy of the wave than in the original noisy signal. Of course this threshold value
functions taken into account in the expansion (4) to reconstruct should eventually depend on the level of noise. The adaptative
the signal or image. Fig. 9 illustrates the variation of the SNR transform makes the filtering of high frequencies stronger at
while reconstructing the denoised result corresponding to the high values of the potential, but the overall level of filtering
signal in Fig. 2(a) (contaminated by additive Gaussian noise should still depend on the noise properties.
of 15 dB) for changing values of the hyperparameter s. For Numerical experiments on the synthetic signal and on the
σ 2 = 0, the reconstructed signal has a SNR worse or similar to cropped Lena in Fig. 2 were carried out to analyze the impact
the original noisy one, indicating once more the importance of of ~2 /2m and σ on the denoising quality and subsequently
 (a) (b)

(c) (d)

(e) (f)

Fig. 12. Influence of the hyperparameters ~2 /2m and σ on proposed decomposition carried out on the 2D system Fig. 2(b) in presence of (a,b) Poisson
noise, (c,d) Gaussian noise and (e,f) speckle noise corresponding to a SNR of 15 dB respectively. The hyperparameters s and ρ have been manually tuned
up to their best possible values for each set of experiment.

to adjust these parameters to their best values for assessment signals were also generated as shown in Fig. 11(d)(g) (Fig.
of the efficiency of the algorithm. Two additional synthetic 11(a) corresponds to the same synthetic signal used previously)
 20

Horizontal distance [pixels]

10
15
20
10

Amplitude
30
5
40

0 50

-5 60

10 20 30 40 50 60
0 100 200 300 400 500
Horizontal distance [pixels]

(a) (b)

(c) (d) (e) (f)

Fig. 13. Signal and images used to compare the proposed denoising method to existing algorithms: (a) Synthetic signal, (b) Synthetic image, (c) Elaine, (d)
Lena, (e) Still image, (f) Moon.

TABLE I
Q UANTITATIVE DENOISING RESULTS .

Gaussian Noise (15dB) Poisson Noise (15dB) Speckle Noise (15dB)

Data Method SNR (dB) PSNR (dB) SSIM SNR (dB) PSNR (dB) SSIM SNR (dB) PSNR (dB) SSIM
Wavelet hard 18.84 25.21 NA 17.21 24.64 NA 17.36 24.13 NA
Wavelet soft 18.53 24.21 NA 17.79 23.62 NA 17.02 22.70 NA
VST NA NA NA NA NA NA NA NA NA
Synthetic Signal
TV 16.20 23.01 NA 15.94 23.45 NA 15.92 22.92 NA
DL NA NA NA NA NA NA NA NA NA
Proposed 22.21 27.50 NA 22.51 27.63 NA 20.75 26.86 NA
Wavelet hard 15.01 24.46 0.61 15.01 25.68 0.69 15.01 25.34 0.76
Wavelet soft 15.71 25.05 0.64 15.61 26.20 0.70 15.49 25.80 0.77
VST NA NA NA 15.09 25.83 0.69 15.06 25.58 0.76
Synthetic Image
TV 15.74 25.07 0.64 15.62 26.23 0.71 15.53 25.78 0.77
DL 17.35 26.15 0.71 17.14 27.22 0.75 17.21 27.48 0.80
Proposed 23.42 31.78 0.89 23.92 32.78 0.92 25.32 33.50 0.95
Wavelet hard 20.52 27.02 0.52 20.08 28.17 0.49 19.75 25.99 0.48
Wavelet soft 21.99 27.69 0.53 21.67 28.59 0.51 21.31 26.61 0.50
VST NA NA NA 21.71 28.64 0.53 22.51 27.81 0.56
Elaine
TV 23.67 29.63 0.62 22.03 28.84 0.55 23.06 27.61 0.59
DL 24.97 29.92 0.68 23.96 29.84 0.62 22.99 27.58 0.58
Proposed 24.70 29.87 0.68 23.89 29.03 0.65 23.52 28.32 0.64
Wavelet hard 20.84 28.17 0.72 20.01 28.89 0.68 19.22 27.49 0.66
Wavelet soft 21.23 28.12 0.71 20.75 28.54 0.67 20.29 27.31 0.66
VST NA NA NA 20.82 29.50 0.73 21.24 28.55 0.69
Lena
TV 21.95 29.32 0.70 21.34 29.58 0.68 21.83 28.71 0.72
DL 23.14 30.02 0.77 21.89 29.61 0.71 20.35 27.24 0.71
Proposed 23.01 29.89 0.78 22.86 29.95 0.77 23.21 30.10 0.78
Wavelet hard 16.89 23.75 0.56 16.99 23.84 0.57 16.82 23.17 0.57
Wavelet soft 17.24 23.64 0.67 17.27 23.63 0.68 17.05 23.04 0.67
VST NA NA NA 19.37 25.36 0.76 19.01 25.01 0.76
Still Image
TV 18.75 24.74 0.75 18.70 24.71 0.75 18.35 24.34 0.74
DL 20.96 26.97 0.78 19.87 27.16 0.71 19.09 25.08 0.68
Proposed 21.39 28.07 0.69 21.93 28.31 0.79 21.83 28.29 0.82
Wavelet hard 20.08 28.89 0.41 19.45 28.48 0.43 18.89 27.03 0.42
Wavelet soft 20.79 29.44 0.45 21.14 29.51 0.77 19.47 27.36 0.75
VST NA NA NA 22.08 30.17 0.85 20.90 28.79 0.84
Moon
TV 21.35 30.22 0.50 23.51 31.41 0.86 21.29 29.18 0.80
DL 22.48 30.84 0.55 22.29 30.40 0.85 19.45 27.59 0.74
Proposed 24.81 33.11 0.73 24.65 33.34 0.86 23.48 31.55 0.89
with the objective of analyzing the behavior of the hyper- value which can be chosen independently of the signal.
parameters for different signal configurations. Three different Finally, Fig. 12 regroups the results for the cropped Lena
types of noise were considered: Poisson, additive Gaussian and image for the three types of noise. The same conclusions can
multiplicative speckle noise. In all cases, the level of noise was be drawn as from the results on 1D signals in Fig. 10: as
adjusted to correspond to a SNR of 15 dB. expected and similar to any other denoising method, the choice
Fig. 10 show the quality of the denoising results for the of the hyperparameters does have an impact on the results, and
initial synthetic signal, in terms of SNR, versus the value the optimal range of parameters depend on the noise. However,
of the hyperparameters ~2 /2m and σ 2 for different types even though the acceptable range of parameters seems smaller
of noise: Poisson noise, Gaussian noise and speckle noise. than for the 1D signal, there is still a relatively large parameter
Several observations can be made from these results. As region where the denoising is very effective. This again makes
expected, an optimal value arises in each case. However, a realistic the possibility to set these parameters beforehand
small variation in the choice of the hyperparameters around in the algorithm independently from the signal or image.
this optimal values only slightly influences the quality of Additionnally, there is a large overlap between the optimal
the denoising. Moreover, the optimal values are only slightly parameter ranges for Poisson and speckle noise, with a marked
dependent on the nature of the noise. This means that for this difference for Gaussian noise. This seems to indicate that the
type of signal the hyperparameters could be fixed beforehand choice of the parameters may differ according to the broad
at a fixed value which can be chosen independently of the type class to which the noise of interest belongs, an information
of noise present. that is usually known beforehand in many cases.
Next, the dependence of ~2 /2m and σ hyperparameters
on the shape of the signals is analyzed. For this purpose, C. Efficiency of the denoising process
three different synthetic signals were generated, as shown in This section presents denoising results on a synthetic signal,
Fig. 11(a)(d)(g), further normalized to 1 and corrupted by a synthetic image and four standard testing images of size
Poisson noise. From the results in Fig. 11(b-c)(e-f)(h-i), it 512 × 512 and 320 × 320 pixels, shown in Fig. 13.
can be clearly observed that the quality of the denoising does Denoising is an extensively explored research field that
depend on the shape of the signals, which can be expected prevents an exhaustive comparison of the proposed approach
given the nature of the adaptative basis used by the proposed to all the existing methods. Moreover, we remind that the
approach. However, the denoising process is efficient for most important contribution herein is to investigate a novel
a fairly large interval around the optimal values. As there way of decomposing signals or images, which is not meant
is a big overlap in the acceptable range of values of the to outperform all the denoising algorithms in any scenario.
hyperparameters for various signal shape, again this means Four algorithms from the literature were used for compar-
that the hyperparameters could be fixed beforehand at a fixed ison purpose: i) wavelet denoising based on hard and soft

(a) (b) (c) (d)

(e) (f) (g)

Fig. 14. (a) Clean Still image, (b) Image corrupted with Gaussian noise corresponding to a SNR of 15 dB. Denoising results obtained using, (c) wavelet
hard thresholding, (d) wavelet soft thresholding, (e) total variation regularization, (f) dictionary learning and (g) proposed method. The proposed adaptive
transform was computed with the hyperparameter ~2 /2m = 0.23, σ 2 = 7.5 and the hyperparameters s and ρ have been manually tuned up to their best
possible values.
 (a) (b) (c) (d)

(e) (f) (g) (h)

Fig. 15. (a) Clean moon image, (b) Image corrupted with Poisson noise corresponding to a SNR of 15 dB. Denoising results obtained using, (c) wavelet
hard thresholding, (d) wavelet soft thresholding, (e) variance stabilization transform, (f) total variation regularization, (g) dictionary learning and (h) proposed
method. The proposed adaptive transform was computed with the hyperparameter ~2 /2m = 0.32, σ 2 = 2.5 and the hyperparameters s and ρ have been
manually tuned up to their best possible values.

(a) (b) (c) (d)

(e) (f) (g) (h)

Fig. 16. (a) Clean Lena image, (b) Image corrupted with speckle noise corresponding to a SNR of 15 dB. Denoising results obtained using, (c) wavelet
hard thresholding, (d) wavelet soft thresholding, (e) variance stabilization transform, (f) total variation regularization, (g) dictionary learning and (h) proposed
method. The proposed adaptive transform was computed with the hyperparameter ~2 /2m = 0.36, σ 2 = 1.35 and the hyperparameters s and ρ have been
manually tuned up to their best possible values.

thresholding of detail coefficients [1], [2], ii) the variance [17], and iv) a dictionary learning (DL) method exploiting
stabilization transform (VST) relevant for data dependent noise sparse and redundant representations over learned patch-based
models [15], iii) an optimization-based approach using the dictionaries [18]. Note that for all the methods and for all
total variation (TV) semi-norm to regularize the solution [16], the simulation scenarios, their hyperparameters were manually
tuned to obtain optimal denoising results in the sense of the SNR. In [21], the quality of CBCT dental image within
quantitative measurements employed. We used the Matlab phantom and in vivo data were evaluated. Fig. 17 shows a
implementations available in the Numerical tours website [19]. noisy image resulting from that study, as well as the denoised
Three quantitative measurements were used to evaluate the images with the proposed approach. The region of interest in
denoised images: the signal to noise ratio (SNR), the peak this image is the dark region in the middle of the tooth, that
signal to noise ratios (PSNR) and the structure similarity represents the canal root. The results displayed show that the
(SSIM) [20]. All the quantitative results are regrouped in method has some practical applications in this field.
Table I where the best and the second best values have
been highlighted by red and blue colors respectively for each IV. C ONCLUSIONS
dataset. Note that VST is only used for data-dependent noise
We investigated in this paper an original approach of
and VST and DL were only tested for images, as initially
constructing an adaptive transform in the context of signal
suggested by the seminal papers. Illustrative results for Still
and image processing based on the resolution of a quantum
image (Fig. 13 (e)) corrupted by Gaussian noise, Moon image
mechanical problem. More precisely, the signal or image is
(Fig. 13(f)) with Poisson noise and Lena image (Fig. 13 (d))
used as the potential in a quantum problem, the resolution of
with speckle noise are shown respectively in Fig. 14, 15 and
which gives as eigenvectors the proposed adaptive basis. The
16. All these results allow us to draw some conclusions. First,
basis vectors automatically use a different range of frequen-
one may remark that in almost all the cases, regardless of
cies to explore low potential valued regions compare to the
the noise nature and the image, the proposed method is one
regions corresponding to the high potential values. Therefore,
of the two best ones. This proves its adaptability to different
thresholding the coefficients of the signal or image expanded
scenarios which can be considered a strong point in number
in this basis will process differently high and low values
of practical applications. Second, we may remark that for
of the signal or image. This framework has been illustrated
the synthetic signal and image, our method outperforms all
through denoising applications on different signals and images
the others. The main reason is that the synthetic signal and
in presence of Gaussian, Poisson and speckle noise. We have
image were generated to promote the main characteristic of
performed a detailed investigation of the impact of the hyper-
the proposed decomposition, that keeps preferentially higher
parameters. We have also presented a quantitative comparison
frequencies for low gray levels and lower frequencies for high
of the denoising efficiency of the proposed adaptive method
gray levels. For such images or signals, the proposed method
compared to state-of-the-art methods on synthetic signals and
is very efficient. On the contrary, TV and DL, for example, fail
standard images. The results of our investigation show that our
in these cases because of the non piece-wise constant nature
method has interesting potential to denoise signals and images,
of the synthetic data. Finally, we remark that the proposed
especially for Poisson and speckle noise to which it is well
denoising algorithm provides competitive results compared to
adapted; indeed, as a vector in the adaptative basis naturally
DL that learns the redundant dictionary from a database of
uses higher frequencies for low values of the signal compared
clean images. Of course the proposed method does not need
to low values, the thresholding process keeps more frequencies
such a database.
for low values than for high values. Our results show that our
denoising procedure outperforms standard methods in specific
D. Application to CBCT dental image denoising cases, and ranks among the best methods in most cases. Our
This section illustrates the ability of the proposed method study of the hyperparameters shows that they cannot be chosen
to denoise real medical images. In particular, the application at random, but that the range of optimality is large enough to
considered in this work for illustration purpose is CBCT allow to set them beforehand independently of the signal or
dental imaging. CBCT is a medical imaging modality that image, although the choice may be modified according to the
allows tooth visualization with low radiation doses, and is thus type of noise present in the application.
suitable for dental applications. However, the low radiation The computational time of the eigenvectors of the Hamil-
prevents the current scanners to provide images with high tonian operator is the major drawback of this method, which

(a) (b) (c)

Fig. 17. (a) Clean CBCT dental image, (b) Noisy CBCT dental image, (c) CBCT dental image after denoising considering the hyperparameter ~2 /2m = 0.5,
σ 2 = 20, ρ = 1 and the hyperparameter s has been manually tuned up to their best possible values.
can be tackled by more refined algorithms. It should be also For a 1D signal, on obtain:
noted that in many applications the computational efficiency Numerical derivatives of ψ: (Dψ)i = ψ(i + 1) − ψ(i)
of the algorithm, while important, is less crucial than the effi- Numerical Laplacian of ψ: (∇2 ψ)i = ψ(i + 1) − 2ψ(i) +
ciency to denoise the signal or image considered. As a future ψ(i − 1)
~2
perspective of this study, it would be very interesting to extend  i = − 2m (ψ(i+1)−2ψ(i)+ψ(i−1))+x(i)ψ(i)
Thus, (Hψ) 
this framework to other reconstruction applications available ~2 ~2 ~2
(Hψ)i = x(i) + 2 2m ψ(i) − 2m (ψ(i + 1) − 2m ψ(i − 1))
in the literature, such as deconvolution, super-resolution or Pi+1
∴ (Hψ)i = j=i−1 Hij ψ(j), for i = 1, 2, 3, ...., N
compressed sensing.
where,

V. ACKNOWLEDGMENTS  ~2
 x(i) + 2 2m f or j = i,
We thank Raphael Smith who participated in a preliminary H(i, j) = ~2
− 2m f or j = i ± 1,
version of this work. We also thank CNRS for funding through 
0 otherwise.
the 80 prime program.
~2 ~2
 
x(1) + 2 2m − 2m
A PPENDIX A ~2
 − 2m
 

E XPLICIT FORM OF H AMILTONIAN MATRIX 



Following equation (1), ⇒H=
 .. .. .. 
~2 . . . 
H = − 2m ∇2 + x.
 
 
Note that:
 
 2 
 − 2m
~ 
∗ The potential V is represented by x (the signal or 2
~2
image). − 2m
~
x(N ) + 2 2m
∗ If x is a signal of size N , then the size of H is For a 2D image x ∈ RN ×N the methodology is similar. A
2 2
N × N. discretized Hamiltonian matrix H ∈ RN ×N is constructed
∗ If x is an image of size N × N , it is transformed using (2), where zero padding is used to handle the boundary
into a vector (in the lexicographical order) of size effects. Precisely if N = 4, i.e. for an image of size 4 × 4
N 2 and H is a N 2 × N 2 matrix. the discretized Hamiltonian becomes of size 16 × 16. The first
∗ In both cases (x is a signal or an image), x is and last six columns and rows of the Hamiltonian matrix are
considered in a vector form. explicitly given in Tables II and III accordingly.

TABLE II
C OLUMN AND ROW 1-6 OF THE H AMILTONIAN MATRIX OF SIZE 16 × 16 CORRESPONDING TO AN IMAGE OF SIZE 4 × 4.

1 2 3 4 5 6
~2 ~2 ~2
x(1, 1) + 2 − 0 0 − 0 ·· ·
2m 2m 2m
~2 ~2 ~2 ~2
− x(2, 2) + 3 − 0 0 − ·· ·
2m 2m 2m 2m
~2 ~2 ~2
0 − x(3, 3) + 3 − 0 0 ·· ·
2m 2m 2m
~2 ~2
0 0 − x(4, 4) + 2 0 0 ·· ·
2m 2m
~2 ~2 ~2
− 0 0 0 x(5, 5) + 3 − ·· ·
2m 2m 2m
~2 ~2 ~2
0 − 0 0 − x(6, 6) + 4 ·· ·
2m 2m 2m

TABLE III
C OLUMN AND ROW 11-16 OF THE H AMILTONIAN MATRIX OF SIZE 16 × 16 CORRESPONDING TO AN IMAGE OF SIZE 4 × 4.

11 12 13 14 15 16
~2 ~2 ~2
·· · x(11, 11) + 4 − 0 0 − 0
2m 2m 2m
~2 ~2 ~2
·· · − x(12, 12) + 3 0 0 0 −
2m 2m 2m
~2 ~2
·· · 0 0 x(13, 13) + 2 − 0 0
2m 2m
~2 ~2 ~2
·· · 0 0 − x(14, 14) + 3 − 0
2m 2m 2m
~2 ~2 ~2 ~2
·· · − 0 0 − x(15, 15) + 3 −
2m 2m 2m 2m
~2 ~2 ~2
·· · 0 − 0 0 − x(16, 16) + 2
2m 2m 2m
 R EFERENCES
[1] David L Donoho and Jain M Johnstone, “Ideal spatial adaptation by
wavelet shrinkage,” biometrika, vol. 81, no. 3, pp. 425–455, 1994.
[2] David L Donoho, Iain M Johnstone, Gérard Kerkyacharian, and Do-
minique Picard, “Wavelet shrinkage: asymptopia?,” Journal of the Royal
Statistical Society: Series B (Methodological), vol. 57, no. 2, pp. 301–
337, 1995.
[3] M. Aharon, M. Elad, and A. Bruckstein, “rmk-svd: An algorithm for
designing overcomplete dictionaries for sparse representation,” IEEE
Transactions on Signal Processing, vol. 54, no. 11, pp. 4311–4322, Nov
2006.
[4] M. Elad and M. Aharon, “Image denoising via sparse and redundant
representations over learned dictionaries,” IEEE Transactions on Image
Processing, vol. 15, no. 12, pp. 3736–3745, Dec 2006.
[5] Raphael Smith, Adrian Basarab, Bertr Georgeot, and Denis Kouamé,
“Adaptive transform via quantum signal processing: application to signal
and image denoising,” in 2018 25th IEEE International Conference on
Image Processing (ICIP). IEEE, 2018, pp. 1523–1527.
[6] Yonina C Eldar and Alan V Oppenheim, “Quantum signal processing,”
IEEE Signal Processing Magazine, vol. 19, no. 6, pp. 12–32, 2002.
[7] C. Aytekin, S. Kiranyaz, and M. Gabbouj, “Quantum mechanics in
computer vision: Automatic object extraction,” in IEEE International
Conference on Image Processing, Sept 2013, pp. 2489–2493.
[8] Taous-Meriem Laleg-Kirati, Emmanuelle Crépeau, and Michel Sorine,
“Semi-classical signal analysis,” Mathematics of Control, Signals, and
Systems, vol. 25, no. 1, pp. 37–61, Mar 2013.
[9] Taous-Meriem Laleg-Kirati, Jiayu Zhang, Eric Achten, and Hacene
Serrai, “Spectral data de-noising using semi-classical signal analysis:
application to localized mrs,” NMR in Biomedicine, vol. 29, no. 10, pp.
1477–1485, 2016.
[10] Abdullah M Iliyasu, “Towards realising secure and efficient image and
video processing applications on quantum computers,” Entropy, vol. 15,
no. 8, pp. 2874–2974, 2013.
[11] Yi Zhang, Kai Lu, Yinghui Gao, and Mo Wang, “Neqr: a novel enhanced
quantum representation of digital images,” Quantum Information Pro-
cessing, vol. 12, no. 8, pp. 2833–2860, 2013.
[12] Zineb Kaisserli and Taous-Meriem Laleg-Kirati, “Image representation
and denoising using squared eigenfunctions of schrodinger operator,”
arXiv preprint arXiv:1409.3720, 2014.
[13] Abderrazak Chahid, Hacene Serrai, Eric Achten, and Taous-Meriem
Laleg-Kirati, “A new roi-based performance evaluation method for
image denoising using the squared eigenfunctions of the schrödinger
operator,” in 2018 40th Annual International Conference of the IEEE
Engineering in Medicine and Biology Society (EMBC). IEEE, 2018, pp.
5579–5582.
[14] P.W. Anderson, “Absence of diffusion in certain random lattices,”
Physical Review, vol. 109, pp. 1492–1505, 1958.
[15] Markku Makitalo and Alessandro Foi, “A closed-form approximation
of the exact unbiased inverse of the anscombe variance-stabilizing
transformation,” IEEE transactions on image processing, vol. 20, no. 9,
pp. 2697–2698, 2011.
[16] M. A. T. Figueiredo and J. M. Bioucas-Dias, “Restoration of poissonian
images using alternating direction optimization,” IEEE Transactions on
Image Processing, vol. 19, no. 12, pp. 3133–3145, Dec 2010.
[17] Leonid I Rudin and Stanley Osher, “Total variation based image restora-
tion with free local constraints,” in Proceedings of 1st International
Conference on Image Processing. IEEE, 1994, vol. 1, pp. 31–35.
[18] Michael Elad and Michal Aharon, “Image denoising via sparse and
redundant representations over learned dictionaries,” IEEE Transactions
on Image processing, vol. 15, no. 12, pp. 3736–3745, 2006.
[19] Gabriel Peyr, “The numerical tours of signal processing,” Computing
in Science & Engineering, vol. 13, no. 4, pp. 94–97, 2011.
[20] Zhou Wang, A. C. Bovik, H. R. Sheikh, and E. P. Simoncelli, “Image
quality assessment: from error visibility to structural similarity,” IEEE
Transactions on Image Processing, vol. 13, no. 4, pp. 600–612, April
2004.
[21] Jérôme Michetti, Adrian Basarab, Michel Tran, Franck Diemer, and
Denis Kouamé, “Cone-beam computed tomography contrast validation
of an artificial periodontal phantom for use in endodontics,” in 2015 37th
Annual International Conference of the IEEE Engineering in Medicine
and Biology Society (EMBC). IEEE, 2015, pp. 7905–7908.