Deep Feature Learning for Wireless Spectrum Data

Abstract

In recent years, the traditional feature engineering process for training machine learning models is being automated by the feature extraction layers integrated in deep learning archi-tectures. In wireless networks, many studies were conducted in automatic learning of feature representations for domain-related challenges. However, most of the existing works assume some supervision along the learning process by using labels to optimize the model. In this paper, we investigate an approach to learning feature representations for wireless transmission clustering in a completely unsupervised manner, i.e. requiring no labels in the process. We propose a model based on convolutional neural networks that automatically learns a reduced dimensionality representation of the input data with 99.3% less components compared to a baseline principal component analysis (PCA). We show that the automatic representation learning is able to extract fine-grained clusters containing the shapes of the wireless transmission bursts, while the baseline enables only general separability of the data based on the background noise.

Full text

Deep Feature Learning for Wireless Spectrum Data
Ljupcho Milosheski, Gregor Cerar, Blaˇ
z Bertalaniˇ
c, Carolina Fortuna and Mihael Mohorˇ
ciˇ
c
Jozef Stefan Institute, Ljubljana, Slovenia
{ljupcho.milosheski, gregor.cerar, blaz.bertalanic, carolina.fortuna, miha.mohorcic}@ijs.si
Abstract—In recent years, the traditional feature engineering
process for training machine learning models is being automated
by the feature extraction layers integrated in deep learning archi-
tectures. In wireless networks, many studies were conducted in
automatic learning of feature representations for domain-related
challenges. However, most of the existing works assume some
supervision along the learning process by using labels to optimize
the model. In this paper, we investigate an approach to learning
feature representations for wireless transmission clustering in
a completely unsupervised manner, i.e. requiring no labels in
the process. We propose a model based on convolutional neural
networks that automatically learns a reduced dimensionality
representation of the input data with 99.3% less components
compared to a baseline principal component analysis (PCA).
We show that the automatic representation learning is able to
extract fine-grained clusters containing the shapes of the wireless
transmission bursts, while the baseline enables only general
separability of the data based on the background noise.
Index Terms—spectrum, analysis, features extraction, self-
supervised, machine learning
I. INTRODUCTION
The introduction of machine learning (ML) algorithms
in wireless communication has led to improvement of the
existing and development of completely new solutions when
sufficient data is available. Some examples include modulation
classification [1], [2], radio technology classification [3], [4],
anomaly detection [5], and device fingerprinting [6], [7]. ML
techniques that rely on manually engineered features from
the data are gradually being replaced by deep learning (DL)
algorithms which are able to extract more relevant features as
an integral part of their training process [8]. Features extracted
using deep-learning models appear to contain more meaningful
information [3] and allow scaling to larger datasets while at
the same time improving the accuracy [7].
Although these DL models provide unmatched accuracy in
domain-based classification tasks, they require large amounts
of labeled data for training, i.e. larger than classical machine
learning algorithms. Such large amounts of training data for
instance on radio spectrum usage are typically collected and
made available for the research community from real-world
environment either using wireless testbed networks such as
LOG-a-TEC [9] or crowd-sourcing initiative such as Elec-
troSense [10]. However, labeling radio spectrum data requires
domain specialists with good knowledge of the operating envi-
ronment and understanding of wireless technologies, making
it an expensive and erroneous process. To address this issue,
usage of unsupervised/semi-supervised [5], [11] models is
emerging as an alternative, but still under-explored approach.
In this paper, we adapt and propose an architecture for
learning feature representations of wireless transmissions from
spectrograms in a completely unsupervised manner. In the
absence of similar approach for direct comparison, we use the
principal components analysis (PCA) as a baseline automatic
representation learning approach. The proposed architecture
was originally developed for feature learning from color
images, known as DeepCluster [12]. We adapt this architecture
to the domain environment and prove it is a worthy alternative
for training a model that outperforms the baseline in the
extraction of features that describe and distinguish the spectro-
gram patterns of different wireless transmission technologies.
Considering the lower amount of content dynamics of the
spectrograms compared to the color images, we propose a
methodology for selecting the number of dimensions that
contain the relevant features in the representation provided by
convolutional neural networks (CNNs).
The main contributions of this work are as follows:
•We propose a CNN-based model that automatically learns
a reduced dimensionality representation of the input data
with 99.5% less components compared to baseline PCA.
•We prove that the proposed CNN-based representation
learning is able to extract features that are representing
actual transmissions, while PCA can learn only general
representations that characterize the background noise.
•We develop a methodology for evaluating the quality
of the provided features with regards to their clustering
tendency, complementary to the clustering quality assess-
ment. The introduction of such evaluation offers addi-
tional insight for the selection of the number of clusters
and number of dimensions of the reduced feature space,
the two critical parameters of the proposed architecture.
The rest of the paper is structured as follows. Section
II analyzes the related work. Section III elaborates on the
feature representation learning using DL while Section IV
elaborates on the experimental methodology, including the fea-
ture development and evaluation metrics. Section V presents
and discusses the experimental results. Finally, Section VI
concludes the paper.
II. RE LATE D WO RK
Regarding the state of the art feature representation learning
approaches available in wireless communications, we identi-
fied two related lines of work: supervised feature learning and
feature learning incorporating unsupervised architectures. The
later can be completely unsupervised or semi-supervised.

A. Supervised feature learning
Considering the amount of research works, supervised DL
architectures have well established usage for the domain-
related problems. For device fingerprinting, as one of the main
tasks in the domain, high classification accuracy is achieved
(above 92%) in [6], [7]. The ability of the CNN to encode
relevant features is also proven in modulation classification
tasks [1], [2] where various types of CNN-based architectures
are used. Supervised solutions for wireless technology classi-
fication achieving high accuracy are proposed in [3], [4].
It is clear that when classification problem is being ad-
dressed and big amount of labeled data is available, CNN-
based solutions achieve top performance. But, providing big
labeled spectrum data, as was discussed before, is an expensive
and erroneous task.
B. Unsupervised and semi-supervised feature learning
General unavailability of labelled spectrum data is con-
straining the usage of the supervised approaches. Thus, efforts
are invested in resolving this problem by using architectures
that require only small section of the data to be labeled,
compromising the accuracy.
In [13], dilated causal convolutional (DCC) architecture is
used in an unsupervised auto-encoder configuration to learn
features from an unlabeled dataset. Small part of the data is
labeled and used for tuning the last layers of the network
in supervised configuration. They show that the auto-encoder
successfully learns the general features of the data.
In [11], an auto-encoder is compared to semi-supervised
bootstrapping of sparse representation for modulation classi-
fication problem. Authors visually show that using the semi-
supervised approach provides better features compared to the
unsupervised approach and generalizes better on unseen data,
but no quantitative support is provided.
There are also usages with completely unsupervised imple-
mentation. In [5], automatic feature learning is proposed with
an auto-encoder network for the task of anomaly detection in
spectrum data. The network is compared with linear and robust
PCA and is shown to better extract features and provides better
accuracy of anomaly detection. But this is still a marginal case
because the task is a binary classification.
In our work, we aim towards completely automatic rep-
resentation learning from large amount of unlabeled radio
spectrum data for the purpose of clustering, when multiple
types of spectrum activities are existing.
III. FEATU RE REPRESENTATION LEARNING
We propose a CNN-based feature learning and clustering
architecture as depicted in Figure 1. It was inspired by
and adapted from the existing DeepCluster model, originally
proposed in [12] for RGB image features learning. The archi-
tecture design contains a representation learning block and a
clustering block. The representation learning contains a CNN
block followed by PCA performing automatic learning of
reduced dimensionality feature representation. The clustering
block then processes the data provided by the representation
learning block. It is relevant to consider that the same dimen-
sionality reduction could be achieved by using one additional
fully connected layer (FCN) after the CNN. However, we use
PCA because it allows automatic feature ranking based on the
explained variance ratio (EVR) as an integral part of the PCA.
The ranking is made in the same Cartesian space in which the
K-means is working, allowing for better explainability of the
developed models.
Compared to the original DeepCluster model [12], we made
the following adaptations:
•Rather than using VGG [14] as a DL architecture, we
selected ResNet18 [15] motivated by the performance
improvement in a use case involving spectrum data in [1].
Thus, we achieve similar performance while reducing the
complexity of the models.
•We customized ResNet input and output layers according
to the shape of the images and the number of classes. In
our case, the input spectrogram images have only one
channel, while the ResNet was originally designed for
3-channel RGB images.
During the training process of the automatic feature repre-
sentation learning using the architecture depicted in Figure 1,
a feedback loop is used as shown with a dotted line. It
consists of a fully connected classification layer attached to
CNN. This layer generates estimated labels used as a ground
truth. During the iterative training process they are compared
to the pseudo-labels generated by the K-means clustering
and the difference is propagated back to guide the training.
More specifically, clustering and CNN weights training are
performed in an alternating manner. In the initial phase, the
K-means clustering on the output of randomly initialized
CNN provides initial cluster assignments which are used as
temporary pseudo-labels (L) for the first epoch of training
of the CNN. The improved CNN is then used for features
extraction in the next iteration, that subsequently with the new
clustering provide new temporal labels. The clustering-training
sequence completes one training iteration. The procedure stops
when the predefined number of iterations (training epochs) is
reached.
It is important to note that through its iterative training
the proposed representation learning approach includes a tight
coupling between the values of the CNN weights, the size
Nof the PCA components and the number of clusters Kas
summarized in the first line of Table I. Using this architecture,
all these dimensions need to be optimized simultaneously and
they influence each other through the feedback loop. However,
in the final application, only the optimized system consisting
of the representation learning is needed. Two possible ways
of utilizing the developed representation learning model are:
1) Feature extractor for clustering which provides ability to
discover new devices, by varying the parameters of the
used clustering algorithm.
2) Transmission classification for already discovered num-
ber of classes using fully connected layer at the output
of the CNN.

Fig. 1: Architecture of the automatic feature learning system.
TABLE I: Adaptable parameters for the Baseline and CNN-
based learning architectures.
Architecture
Function
block Representation learning Clustering
CNN-based CNN + PCA (n=1..N) (k=1..K)
Baseline PCA (n=1..N) k (K=2..30)
The equivalent baseline architecture not involving the CNN
learning component and the dotted training loop in Figure 1
can be optimized in a sequential manner: fist optimizing the
PCA as a representation learning method and then optimizing
the K-means as clustering method. The equivalent optimization
parameters are summarized in the second line of Table I. The
baseline system employs a flattening block that reorders the
elements of the input matrix into a single row and feeds them
to PCA to learn a representation.
IV. METHODOLOGY
A. Training and Evaluation Data
The dataset used for the performance analysis consists of 15
days of radio spectrum measurements acquired in the LOG-
a-TEC testbed at a sampling rate of 5 power spectral density
measurements per second using 1024 FFT bins in the 868 MHz
license-free (shared spectrum) band with a 192 kHz bandwidth.
Details of the acquisition process and a subset of data can be
found in [9]. The acquired data has a matrix form of 1024×M,
where Mis the number of measurements over time.
Fig. 2: Sample of 8 spectrogram segments from the data.
The complete data-matrix was segmented into non-
overlapping square images (spectrograms) along time and
frequency (FFT bins) for a window size W= 128. An
example of such segmentation containing 8 square images is
shown in Figure 2, corresponding to the image resolution of
25.6 seconds (128 measurements taken at 5 measurements per
second) by 24 kHz. The window size is large enough to contain
any single type of activity and small enough to avoid having
too many activities in a single image while also having in
mind computational cost. Dividing the entire dataset of 15
days using W= 128 and zero overlapping, produces 423,904
images of 128 ×128 pixels. Additionally, the pixel values are
scaled to [0,1].
B. Optimization of the Representation Learning
CNN-based and baseline approaches can be optimized along
two dimensions: the number of PCA components that should
be used in the representation and the number of clusters for
the K-means. For the baseline model, the two parameters are
independent, meaning that the representation learning function
is not affected by the number of clusters that will be later used
on the obtained feature vectors. On the other side, for the
CNN-based architecture, changes in the number of clusters
affect the representation learning. This is because the number
of clusters should always be the same as the number of classes
at the output of the CNN during the learning process, so
it inflicts changes on the representation learning block. This
means that varying the number of clusters should also be
considered when choosing the number of dimensions for the
representation learning with the CNN-based model.
It is unfeasible to study the influence of individual CNN
weights on the learnt representation and cluster quality due
to their large numbers. They are optimized in a black-box
manner during the training process consisting of 200 training
epochs. This number was determined empirically by observing
the convergence of the loss function.
C. Evaluation
As an evaluation metric for choosing the dimensionality of
the representation for both models we use EVR [16]. EVR is
a measure of how much of the variation in the feature space is
assigned to each of the principal components after performing
PCA.
We analyze the quality of the representation for clustering
purposes employing visual assessment of tendency (VAT) [17].
This method produces matrix visualisation of the dissimilarity
of randomly selected subset of samples based on their pairwise

euclidean distances. The samples are ordered in such a way
that groups that are closely located in the feature space,
according to the distance metric, appear as dark squares along
the diagonal of the matrix. Implementation wise, we used an
improved version of VAT (i.e., iVAT) which provides better
visualization than the standard one.
We also evaluate the quality of the clustering, performed on
the extracted features, by using the Silhouette score metric. In
this way we provide quantification of how well the clusters
are distinguished for the analysed models.
Using these metrics, we evaluate and explore the repre-
sentation learning capabilities of both approaches and their
applicability for clustering. First we analyze a histogram of
samples for the formed clusters based on the frequency sub-
band resulting from image segmentation that each sample
comes from. These plots provide information on whether
the learned feature representation used for the clustering is
correlated to the location of the samples along the frequency
axis. Then we plot the average of the samples assigned to
a single cluster. This provides an insight into the actual
spectrogram content that is specific for the formed clusters.
V. EXPERIMENTAL RESULTS
A. Learning with the PCA baseline approach
In Figure 3, we present the evaluation of the learnt rep-
resentation according to EVR and VAT metrics discussed in
Section IV-C. Figure 3a shows EVR of the features learnt by
the baseline representation learning block consisting of PCA
only, followed by the VAT plots in Figures 3b–h. The plots
correspond to 7 different PCA-based representation learning
models, configured for different number of components se-
lected in a way to evaluate the feature vectors with wide range
of different dimensions.
It can be seen that by keeping 95% of the variance ratio
in the PCA learned representation provides feature vectors of
dimension 1x3770, which is around 23% of the flattened single
sample input of 1x16384. Although the learned representation
has reduced dimensions by more than four times compared
to the flattened input, we still have a high dimensionality
representation. The VAT plots in Figure 3b–h show that the
baseline approach learns representation with weak clustering
tendency for all cases, except the one when using only the
first two components of the feature space. The VAT-2 plot
(Figure 3b) of the 2-dimensional feature representation shows
the existence of three well separated clusters.
B. Learning with the CNN-based approach
Figure 4a shows EVR of the features learnt by the Rep-
resentation learning block using ResNet18 (RN) and VGG11
(VGG) DL-based models with different number of clusters. A
smaller number of clusters yields higher EVR in the lower
components, while there is no significant difference in the
cumulative sum of EVR after the 20th component across
different models. All models encode features with more than
95% of EVR within 27 components which is 0.7% of the 3770
components required by the baseline PCA.
(a) Cumulative sum of EVR for the baseline representation.
(b) VAT-2 (c) VAT-5 (d) VAT-25 (e) VAT-250
(f) VAT-1000 (g) VAT-2000 (h) VAT-3770
Fig. 3: Evaluation of the baseline representation learning.
(a) Cumulative sum of EVR for DL-based representations.
(b) RN-10 (c) RN-20 (d) RN-25 (e) RN-30
(f) VGG-10 (g) VGG-20 (h) VGG-25 (i) VGG-30
Fig. 4: VAT plots for DL-based representations.

The VAT plots for the proposed RN-based model are shown
in Figure 4b–e and for the VGG-based model in Figure 4f–i.
The plots correspond to 4 different automatic representation
learning models, configured for 10, 20, 25 and 30 clusters.
Both DL models achieve very similar EVR. Experimentally
it was observed that models trained with lower number of
clusters significantly worsen the clustering tendency according
to VAT plots, so they were not considered in the subsequent
analysis. For a smaller number of clusters in Figure 4b,
the learnt representations with RN-based model contain less
prominent dark squares compared to the models with larger
number of clusters in Figures 4b–c. Similar observation holds
for VGG-based models in Figures 4e–h. For both DL models,
the 25 clusters models show the most distinguished separation
of the feature space. This analysis indicates that the proposed
architecture is able to learn representations that can yield
11 to 25 well separated clusters. It is also able to learn
5 to 10 and 26 to 30 less clearly separated clusters while
it is less suitable for small number of clusters such as 2
to 4. Overall, the automatically learnt representation is able
to extract fine-grained clusters containing the shapes of the
wireless transmission bursts.
Comparing to the baseline, the CNN-based model can learn
to encode the relevant information for cluster development in
only approximately 0.7% of the components required by the
PCA baseline, when the application requires higher numbers
of well defined clusters, while also enabling superior cluster
differentiation. For two clusters, the baseline model provides
a better separation according to the VAT plots and the same
feature dimensionality size.
C. Cluster analysis
Next we examined the best clusters developed with the
baseline and the CNN-based approach using histograms of
samples accompanied with the average cluster spectrograms
in Figure 5 and Figure 6, respectively.
For the best baseline approach containing 3 clusters, Fig-
ure 5 shows that the cluster 0 contains mostly the samples from
sub-bands between 1 and 7, the cluster 1 contains the samples
from the left-most sub-band and the cluster 2 contains almost
all of the samples from the right-most sub-band. Sub-bands
refer to sections of frequency-wise segmentation as shown in
Figure 2. This observation is also aligned with the size of the
dark squares in the VAT plot of the 2 PCA features in Figure
3b, which contains 3 clusters, one big corresponding to the
cluster 0, and two almost equal smaller clusters corresponding
to the clusters 1 and 2. Clearly, the baseline approach clusters
the data based on the weaker signal on the left-most and right-
most samples of the full bandwidth. The weaker signal seems
to be a consequence of the nonuniform sensing capability
of the sensor. Looking back to Figure 5, the left-most and
the right-most samples have gradually vanishing brightness
towards the edges. Plotting the average of the assigned spec-
trograms from each of the clusters supports this observation.
Experimentally we identified that the 24-cluster automatic
CNN-based model using ResNet18 DL architecture provides
Fig. 5: Distribution of samples from each cluster along the
frequency band for the baseline approach.
Fig. 6: Distribution of samples from each cluster along the
frequency band for the clusters obtained by the RN-based
approach.
the best results. For this model, the average spectrograms and
histograms in Figure 6 show its effectiveness in learning gen-
eral features related both to the transmission-specific content
and the ”background” of the spectrograms. The combined
clusters 19 and 23 in Figure 6 are the same as the cluster
2 from the baseline approach, occupying the right-most sub-
band, while the cluster 7 corresponds to the cluster 1 of
the baseline. According to Figure 6, the samples assigned
to this cluster are again occupying the right-most sub-band.
This means that the automatic model can also learn the
features extracted with the baseline approach. Additionally,
the automatic model learns features that are specific for the
different patterns generated by the transmissions. The clusters

0, 2, 4, 5, 8, 9..14, 17, 18, 20..22, in Figure 6 show horizontal
line activities, which according to Figure 2 appear across
the entire bandwidth. Their histograms show that samples
assigned to these clusters are from all 8 sub-bands, and their
distribution along the entire channel is roughly uniform. The
clusters 3, 6 and 15 show the capability of the automatic model
to distinguish the transmission-free spectrograms. This can
be used to determine transmission-free sub-bands, which is
another advantage against the baseline approach. Finally, the
clusters 1 and 16 show groups of dot-like transmission bursts.
Fig. 7: Silhouette scores
The Silhouette scores depicted in Figure 7 for the baseline
and the two CNN-based approaches confirm the observations
based on the VAT plots. The baseline approach exhibits better
performance at small number of clusters, for three clusters
achieving the best score of 0.68. However, the baseline-
provided feature space does not show transmission specific
groups as samples are clustered based on the background
noise. The automatic CNN-based models show comparable
performance across the variation of clusters. This justifies the
usage of the proposed lower complexity CNN, the ResNet18
instead the VGG11, preserving the performance and signif-
icantly reducing the complexity in terms of the number of
required DL model parameters by roughly 11 times, according
to Table II.
TABLE II: Complexity comparison.
Algorithm RN VGG Baseline
Num.
parameters
11 M 133 M /
VI. CONCLUSIONS
In this paper, an automatic feature representation learning
architecture based on CNN and PCA was explored and com-
pared to a baseline model using only PCA for the task of
clustering spectrograms from radio spectrum measurements.
Our findings show that the baseline approach is useful when
clustering based on general features of the data is required,
with only a small number of clusters. On the other side, the
automatic learning combining CNN and PCA, although more
complex, provides much finer distinction between closely
related groups of samples, based on their actual content, which
are the transmissions bursts in our case. This shows that such
architecture can be used for automatic representation learning
and is suitable when large yet unlabeled spectrogram data is
available.
ACKNOWLEDGMENTS
This work was funded in part by the Slovenian Research
Agency under the grant P2-0016 and in part by the European
Union’s Horizon Europe Framework Programme under the
grant agreement No 101096456 (NANCY). The project is sup-
ported by the Smart Networks and Services Joint Undertaking
and its members.
REFERENCES
[1] T. J. O’Shea, T. Roy, and T. C. Clancy, “Over-the-air deep learning based
radio signal classification,” IEEE Journal of Selected Topics in Signal
Processing, vol. 12, no. 1, pp. 168–179, 2018.
[2] S. Rajendran, W. Meert, D. Giustiniano, V. Lenders, and S. Pollin, “Deep
learning models for wireless signal classification with distributed low-
cost spectrum sensors,” IEEE Transactions on Cognitive Communica-
tions and Networking, vol. 4, no. 3, pp. 433–445, 2018.
[3] J. Fontaine, E. Fonseca, A. Shahid, M. Kist, L. A. DaSilva, I. Moer-
man, and E. De Poorter, “Towards low-complexity wireless technology
classification across multiple environments,” Ad Hoc Networks, vol. 91,
p. 101881, 2019.
[4] E. Fonseca, J. F. Santos, F. Paisana, and L. A. DaSilva, “Radio
access technology characterisation through object detection,” Computer
Communications, vol. 168, pp. 12–19, 2021.
[5] Q. Feng, Y. Zhang, C. Li, Z. Dou, and J. Wang, “Anomaly detection
of spectrum in wireless communication via deep auto-encoders,” The
Journal of Supercomputing, vol. 73, no. 7, pp. 3161–3178, 2017.
[6] K. Merchant, S. Revay, G. Stantchev, and B. Nousain, “Deep learning
for rf device fingerprinting in cognitive communication networks,” IEEE
Journal of Selected Topics in Signal Processing, vol. 12, no. 1, pp. 160–
167, 2018.
[7] J. Robinson, S. Kuzdeba, J. Stankowicz, and J. M. Carmack, “Dilated
causal convolutional model for rf fingerprinting,” in 10th IEEE Annual
Computing and Communication Workshop and Conference (CCWC),
2020, pp. 0157–0162.
[8] S. Riyaz, K. Sankhe, S. Ioannidis, and K. Chowdhury, “Deep learning
convolutional neural networks for radio identification,” IEEE Communi-
cations Magazine, vol. 56, no. 9, pp. 146–152, 2018.
[9] T. ˇ
Solc, C. Fortuna, and M. Mohorˇ
ciˇ
c, “Low-cost testbed development
and its applications in cognitive radio prototyping,” in Cognitive radio
and networking for heterogeneous wireless networks. Springer, 2015,
pp. 361–405.
[10] S. Rajendran, R. Calvo-Palomino, M. Fuchs, B. Van den Bergh, H. Cor-
dob´
es, D. Giustiniano, S. Pollin, and V. Lenders, “Electrosense: Open
and big spectrum data,” IEEE Communications Magazine, vol. 56, no. 1,
pp. 210–217, 2017.
[11] T. J. O’Shea, N. West, M. Vondal, and T. C. Clancy, “Semi-supervised
radio signal identification,” in 19th IEEE International Conference on
Advanced Communication Technology (ICACT), 2017, pp. 33–38.
[12] M. Caron, P. Bojanowski, A. Joulin, and M. Douze, “Deep clustering
for unsupervised learning of visual features,” in European Conference
on Computer Vision (ECCV), 2018, pp. 132–149.
[13] S. Kuzdeba, J. Robinson, and J. Carmack, “Transfer learning with radio
frequency signals,” in 18th IEEE Annual Consumer Communications &
Networking Conference (CCNC), 2021, pp. 1–9.
[14] K. Simonyan and A. Zisserman, “Very deep convolutional networks for
large-scale image recognition,” arXiv preprint arXiv:1409.1556, 2014.
[15] K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for
image recognition,” in IEEE Conference on computer vision and pattern
recognition, 2016, pp. 770–778.
[16] I. Jolliffe, “Principal component analysis,” Springer: Berlin, Germany,
vol. 87, pp. 41–64, 01 1986.
[17] J. C. Bezdek and R. J. Hathaway, “Vat: A tool for visual assessment of
(cluster) tendency,” in IEEE international Joint Conference on Neural
Networks (IJCNN’02), vol. 3, 2002, pp. 2225–2230.