This article has been accepted for publication in a future issue of this journal, but has not been

fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2019.2930296, IEEE Access

Date of publication xxxx 00, 0000, date of current version xxxx 00, 0000.

Digital Object Identifier 10.1109/ACCESS.2019.Doi Number

Classification of heterogeneity on multi-spectral

transmission image based on modulation-
demodulation-frame accumulation and pattern
recognition
Fulong Liu12, Gang Li12, Wenjuan Yan3, Guoquan He3, Shuqiang Yang4 And Ling Lin12
1
State Key Laboratory of Precision Measuring Technology and Instruments, Tianjin University, Tianjin 300072, China
2
Tianjin Key Laboratory of Biomedical Detecting Techniques and Instruments, Tianjin University, Tianjin 300072, China
3
School of Electronic Information Engineering, Yangtze Normal University, Chongqing 408100, China
4
School of physics and electronic information, Luoyang Normal University, Luoyang 471000, China

Corresponding author: Ling Lin (e-mail: linling@tju.edu.cn).

This article does not have sponsorship or financial support.

ABSTRACT Multi-spectral transmission image provides a possibility for the detection of early breast cancer.
However, in the process of acquiring multi-spectral transmission images, the recognition of heterogeneities has
many difficulties due to the image blur caused by the scattering effect of light source in biological tissues and
weak transmission signals. This paper proposes a combination method of modulation-demodulation-frame
accumulation technique and pattern recognition to achieve heterogeneous classification. Firstly, the acquisition
experiment of the phantom multi-spectral images is designed. Then, the signal-to-noise ratio (SNR) of the image
is improved by the modulation-demodulation and frame accumulation technique, and the 14-dimensional feature
information (firmness, angular second-order distance, contrast, gray-scale correlation, entropy, inverse gap,
smoothness, dissimilarity, consistency, center of gravity, area, perimeter, long diameter of irregular image, short
diameter of irregular image) of the heterogeneous region are extracted from the image with high SNR. Finally, the
heterogeneous classification accuracy of different models is compared. The results show that: Compared with the
classification accuracy of the traditional multi-spectral image classification models, Random Forest (RF) and
Extreme Learning Machine (ELM) models have better classification effect when subdividing the four types of
heterogeneity based on the data set of this paper. Among them, the RF and ELM models established by the dataset
of four-wavelength combination have the best classification effect, and the classification accuracy rate reaches
100%, secondly, it is the three-wavelength combined model. The single-wavelength model has the worst
classification effect. And the operating efficiency of ELM is significantly higher than RF. In conclusion, the image
quality is improved by modulation-demodulation and frame accumulation technique. And compared with the
classification accuracy of the traditional multi-spectral image classification models, the RF and ELM models
established in this paper have better classification effect, which may promote the application of multi-spectral
transmission imaging in early screening of breast tumors.

INDEX TERMS Multi-spectral transmission image, Modulation-demodulation and frame accumulation,

Heterogeneous classification, Pattern recognition

I. INTRODUCTION late stage [3]. So early detection of breast cancer is critical.

Breast cancer is a common disease in adult women, which is a
serious threat to women's health and life [1]. In 2018, the China
Cancer Center announced that the incidence of breast cancer in
Chinese women was 11.2%, and the mortality rate was 9.2%,
ranking first in the world [2]. Early detection of breast cancer
is conducive to breast conserving treatment, and the cure rate
of early stage patients is much higher than that of middle and
Currently, imaging techniques for detecting breast cancer
mainly include x-ray [4-6], ultrasound [7-9], magnetic
resonance imaging (MRI) [10,11], computed tomography (CT)
[12] and near-infrared transmission imaging [13-15]. However,
due to x-ray radiation, it is detrimental to human biological
tissues. Ultrasonic testing has different detection sensitivities
for different populations, and it is difficult to guarantee the

VOLUME XX, 2019 1

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2019.2930296, IEEE Access
Fulong Liu: Classification of heterogeneity on multi-spectral transmission image based on modulation-demodulation-frame accumulation
standard of detection. The cost of MRI testing is too high to be
promoted. In addition, the spatial distribution of early tumors
in these images is ambiguous, making it difficult to identify
heterogeneous regions. Considering the mammary gland is a
transparent body without bone, and the tumor tissue contains a
large number of new blood vessels and hemoglobin, there will
appear a large shadow in the transmission imaging (in the tissue,
uneven medium called heterogeneity [16]). Therefore, optical
transmission imaging provides a feasible and simple method
for early detection of breast cancer.
In optical transmission imaging, multi-spectral non-
destructive optical testing has become a hot topic due to its
advantages of real-time, non-invasive, safe, specific and
sensitive and has been widely used in many fields [17-21], but
there are few studies on the application of multi-spectral
transmission images in the medical field. This is mainly due to
the strong scattering effect of the incident light in the
transmission process of biological tissue, which makes the
signal in the multi-spectral transmission image weak and
unable to obtain the rich characteristic information of
heterogeneities. In recent years, the technique of modulation-
demodulation (loading shaped signals) and frame
accumulation has become one of the most effective methods to
enhance low-light-level (LLL) diffuse reflection image signals.
Among them, Li et al uses the combination method of frame
accumulation and shaping signal technology to greatly enhance
the SNR of the LLL image and improve the resolution of the
image [22-24]. Therefore, this paper, for the first time, attempts
to apply the combination method of modulation-demodulation
and frame accumulation to the transmission images to improve
the image quality.
In addition, in the multi-spectral transmission imaging
process, the optical properties of different tissues at different
wavelengths are different, which is beneficial to the
classification of heterogeneities. Although the spatial structure
of the early tumor tissue is similar to that of the normal tissue,
the characteristic of the above multi-spectral image provides a
good idea for identifying early tumor tissue in the image. And
as a powerful technique to analyze and extract unique
information from multi-spectral image, feature information-
based classification has been an active research topic in recent
years. Many classification algorithms have been developed to
perform multi-spectral image classification. For example,
support vector machines (SVM) [25,26], linear discriminant
analysis (LDA) [27,28], low-rank representation (LRR) [29-
31], sparse representation (SR) [32-34] and multinomial
logistic regression (MLR) [35] are well-known image
classification methods. Among them, Gao et al used the
multiple kernel learning method to fuse different features of
samples to improve the classification accuracy of SVM in
multi-spectral images [36]. Wang et al proposed a novel LDA
to obtain all representative subspaces by adaptively, which
improved the classification accuracy of different objects in
multi-spectral images [37]. Wang et al proposed a local and
structural-regularized LRR for multi-spectral image
classification, which overcomes the limitations of multi-
dimensional image feature extraction [38]. Gao et al improved
SR to improve the precision of classification result by
incorporating the neighboring information of the test pixel [39].
So, multi-spectral transmission image provides a possibility for
2 VOLUME XX, 2019
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
and pattern recognition
the detection of early breast cancer.
In this paper, we combine the method of enhancing the LLL
image signal with the model of pattern recognition to achieve
effective classification of different heterogeneous tissues in
multi-spectral transmission images. Different from the
traditional multi-spectral image classification method
mentioned above, which adaptively extracts image feature
information, we extract the feature information of the
heterogeneities on the different wavelength images from three
aspects of spatial feature, texture feature and shape feature as
the classification basis of the heterogeneities. And considering
the nonlinearity among experimental data, we use the
traditional pattern recognition method to classify the
heterogeneous tissues. The traditional pattern recognition is
divided into supervised classification and unsupervised
classification. Different heterogeneities are used to simulate
tumor masses, and the categories of heterogeneities are known.
Therefore, we mainly study the methods of supervised
classification pattern recognition. The methods commonly
used in supervised pattern recognition include RF with multi-
layer neural networks and ELM with single-layer neural
network. RF, first proposed by Breiman, is widely used in data
processing, text classification and other fields because of its
advantages such as simple implementation, strong anti-
overfitting ability, parallel processing and good ability to deal
with nonlinear modeling problems [40]. In order to solve the
problem that multi-neural networks consume a lot of time in
the process of classification optimization algorithm, Huang
proposed a single hidden layer feed-forward neural network
ELM algorithm, which has the characteristics of simple model
design, fast operation speed and high generalization
performance, and has good performance in multi-label learning
[41]. Finally, the SNR and gray resolution of the images are
improved through the technique of modulation-demodulation
and frame accumulation. And the 14-dimensional feature
information of the heterogeneous region are extracted from the
image with high SNR for pattern recognition. The results show
that compared with the accuracy of traditional multi-spectral
image classification models (SVM, LDA, LRR, SR and MLR),
the RF and ELM we established both have a better
classification effect when subdividing 4 heterogeneities under
the data set of this paper, which promotes the application of
transmission multi-spectral imaging in early screening of
breast cancer.
II. RELATED TECHNOLOGY
A. MODULATION-DEMODULATION AND FRAME
ACCUMULATION TECHNOLOGY
In the process of multi-spectral transmission imaging, the
distribution of light intensity may be affected due to the
influence of the external environment and the scattering effect
of heterogeneities themselves, resulting in weak signal and low
definition of transmission images. Therefore, this paper needs
to perform certain pre-processing experiments on the obtained
transmission images. Among them, the weak signal in the
image includes two meanings: one is that the absolute value of
the signal light intensity is low, and the other is that the SNR
and resolution of the signal are low. The modulation-
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2019.2930296, IEEE Access
Fulong Liu: Classification of heterogeneity on multi-spectral transmission image based on modulation-demodulation-frame accumulation
Heterogeneity
Phantom
Power
Source
Supply
Camera
LED
CCS
Sealed MFSG
Dark Box
(a)
FIGURE 1.
demodulation and frame accumulation technology has become
an effective mean to enhance the LLL image signal.
Gray level is the core of image accuracy and heterogeneity
detection sensitivity. The more grayscale of image, the richer
the image information, the better the classification of
heterogeneous tissue in tissue. Low-resolution images become
obstacles to tissue classification and spatial information
extraction, but the modulation-demodulation and frame
accumulation technique can improve the resolution and
increase the grayscale of images to some extent [22]. The
modulation and demodulation techniques are the process of
moving the spectrum of the baseband signal to the channel
passband to achieve the purpose of enhancing the signal, and
then restoring the band signal brought by the channel back to
the baseband signal [23,42]. The frame accumulation technique
adds the gray values of the corresponding pixel points of the
multi-frame images to obtain the average image in the time
series, filters out the noise in the image, and multiples the SNR
of the image [22]. Therefore, this paper uses the technique of
modulation-demodulation and frame accumulation to improve
the SNR and resolution of the transmission images.
B. TARGET CLASSIFICATION
Pattern recognition, also known as pattern classification, is
divided into supervised classification and unsupervised
classification from the perspectives of problem processing
nature and problem solving method. This paper focuses on the
study of supervised classification methods, mainly comparing
the classification effects of RF with multi-layer neural network
and ELM with single-layer neural network. Simultaneously, in
order to prove the validity of the supervised classification
models (RF and ELM) established in this paper, we compare
them with the traditional multi-spectral image classification
models.
RF is an integrated learning method based on Bagging,
consisting of multiple unpruned decision trees. In a
classification task, the categories of test samples are
determined by the mode of the output category labels of these
decision trees. Each decision tree and out of bag data (OOB
data) in the RF are obtained by bootstrap resampling. The
importance of the characteristic variables of the decision tree is
calculated according to the Gini index [43]. Traditional multi-
VOLUME XX, 2019
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
and pattern recognition
Different shapes
Different components
(b)
Diagram of experimental device.
neural network algorithms need to set more network
parameters. When solving the optimal solution, local optimal
solutions may appear and the global optimal solution of the
network cannot be obtained. In the ELM solution, only the
number of hidden layer nodes needs to be set, and the global
optimal solution can be solved by randomly initializing the
weights and offsets. ELM solves the feedforward neural
network with single hidden layer, which can be divided into
two stages: random feature mapping and linear parameter
solving.
III. EXPERIMENT
According to the characteristics of breast tissue, this paper
designed the collection experiment of phantom. The 14-
dimensional feature information of the different wavelength
images is respectively extracted on the multi-spectral
transmission images after the modulation-demodulation and
frame accumulation processing. And the different
heterogeneities are respectively classified by pattern
recognition methods (RF and ELM) and traditional multi-
spectral image classification models.
A. EXPERIENTAL DEVICE
Fig.1 depicts a schematic of the experimental device. This
experimental device is mainly divided into five parts: light
source part, control part, image acquisition part, processing
part and other parts. The light source part mainly includes: 4
wavelength LEDs (blue wavelength centered at 435nm, green
wavelength centered at 546nm, red wavelength centered at
700nm and near-infrared wavelength centered at 860nm) and
programmable DC regulated power supply (model: hspy-600).
The control part uses the LED drive circuit of PT4115 to
control the input of signal. The image acquisition part mainly
includes: industrial camera (model: JHSM120f, camera frame
rate: 54 frames/s and image resolution: 1280 × 960) and
acquisition software (usbVideo). The processing part mainly
adopts HP computer and MATLAB software to process related
images. The other parts mainly include: blackout cloth and
phantom. Phantom composition: a highly photometric PMMA
(polymethyl methacrylate) rectangular container (the
transmittance of the breast is higher than that of other human
tissues) is used to hold the mixed solution of milk and water
3
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2019.2930296, IEEE Access

Fulong Liu: Classification of heterogeneity on multi-spectral transmission image based on modulation-demodulation-frame accumulation
and pattern recognition

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

FIGURE 2. Phantom multi-spectral transmission images. (a) blue wavelength transmission image; (b) green wavelength transmission image; (c) near-infrared
wavelength transmission image; (d) red wavelength transmission image.

(a) (b)
FIGURE 3. Fourier transform diagram. (a) frequency coordinate position map; (b) frequency map.

FIGURE 4. Cycle diagram of the sinusoidal shaped signal

with different concentrations. According to the characteristics follows:

of strong transmissibility and tomographic distribution of
breast tissue, potatoes and carrots are suspended in the solution
as heterogeneities to simulate breast cancer [44].
B. IMAGE ACQUISITION
The original multi-spectral images are obtained on the
experimental device that had been set up: milk solution +
heterogeneous image. The specific experimental steps are as
(1) Set the parameters of camera and signal generator. The
gain in the camera is set to 10, the exposure time is set to
101s, the acquisition rate is about 45 frames/s. And the
frequency of the sinusoidal shaped signal in the signal
generator is set to 4Hz.
(2) Adjust and fix the distance between the light source and
the phantom, the phantom and the camera separately, and
set up the blackout cloth. Turn on the light source switch

4 VOLUME XX, 2019

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2019.2930296, IEEE Access
Fulong Liu: Classification of heterogeneity on multi-spectral transmission image based on modulation-demodulation-frame accumulation
and the camera of the camera to get a sequence of images
of the phantom.
(3) The four kinds of light sources loaded with the sinusoidal
shaped signal are respectively irradiated with the phantom.
The phantom included four heterogeneities (two potato
pieces and two carrot pieces) of different sizes and
thickness, and the experiment is carried out in groups.
Each wavelength includes 6 groups, each group
configured with different concentrations of solutions, a
total of k  24 groups.
(4) Excluding the images with large errors in each group, a
total of n  26880 original multi-spectral images are
obtained, and Fig.2 are one frame of the original phantom
images x il, j ( l  1, 2,  ,1120; i  1, 2, 3, 4; j  1, 2, ..., 6) of
the four wavelengths, respectively.
C. IMAGE PREPROCESSING
The multi-spectral transmission images obtained in this
experiment are processed by the technique of modulation-
demodulation and frame accumulation. The specific image
processing steps are as follows:
(1) Perform a fast Fourier Transform (FFT) on images
x il, j ( l  1, 2,  ,1120; i  1, 2, 3, 4; j  1, 2, ..., 6) , and extract
the coordinate values corresponding to the frequency
components of the four wavelengths, as shown in Fig.3.
(2) According to the frequency coordinate value determined
above, the pixels of all the images of the four wavelengths
are demodulated, and the demodulated images of all
wavelengths are obtained.
(3) Read the demodulated images x il, j ( l  1, 2,  ,11 20 )
into MATLAB program respectively, find the sum of gray
values of each frame, and draw 1120 gray value of images
xil, j (l  1, 2, 3,...,1120) , as shown in Fig.4.
(4) Determine the number of image frames in a single sine
wave period according to the curve shown in Fig.4. It can
be seen from the figure that each sinusoidal shaped signal
includes 11 frames images.
(5) Perform the frame-accumulate average for every 11
frames of images each group in the experiment, that is:
l 11
x l
i, j
xk
i, j  l
(1)
f image segmentation threshold in the same wavelength is small,
i  1, 2,3, 4; j  1, 2,..., 6; l  1, 2, ,1109; k  1, 2,..., 606; f  11.
(6) According to formula (1), a total of n2  2424 frame
accumulative multi-spectral images are obtained, in
which each wavelength includes 606 images.
D. HETEROGENEOUS CLASSIFICATION
1) FEATURE INFORMATION EXTRACTION
In order to accurately extract the information in each
heterogeneous region, each preprocessed image is reasonably
cropped into four regions of uniform size without affecting the
accuracy of the algorithm and image quality (each region
includes one heterogeneity). Different from the traditional
multi-spectral image classification methods mentioned above,
which adaptively extracts image feature information, the
VOLUME XX, 2019
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
and pattern recognition
relevant information of the image is extracted from the spatial
features, the texture features and shape features.
Spatial feature extraction: The spatial relationship
indicates that there is a certain spatial positional relationship
and directional relationship among multiple objects in the
image, such as the firmness of the image's adjacency
relationship. Texture feature extraction: The gray-level co-
occurrence matrix (GLCM) is used to extract the texture
features of the image, which uses the probability in statistics to
reflect the overall information of the image grayscale related
directions and intervals. There are 8 commonly used texture
feature parameters (angular second-order distance, contrast,
gray-scale correlation, entropy, inverse gap, smoothness,
dissimilarity and consistency). Shape feature extraction:
Hough transform is used to detect the heterogeneous regional
feature information of the image, which mainly includes five
feature parameters (center of gravity, area, perimeter, long
diameter of irregular image, short diameter of irregular image).
Among them, in order to precisely determine the region of the
heterogeneities, the heterogeneous feature region is obtained
by using the Ostu threshold segmentation method for different
wavelength images. In the process of obtaining mask images
by Ostu threshold segmentation method, the coincidence
degrees of different heterogeneous mask images and true
knowledge map (the actual heterogeneous images were
obtained before the experiment) are calculated as the
discrimination basis for the selection of threshold values, so as
to obtain the multi-wavelength mask images with the highest
coincidence degree. Because the mask image needs to cover
the target region of the sample as much as possible, it needs to
have a high degree of coincidence with the actual target region
of the sample to improve the accuracy of heterogeneity
detection. According to comprehensive statistics, the
coincidence degree of all multi-spectral mask images reached
more than 96%, which was enough to cover the characteristic
information of heterogeneity in the images. And the average
threshold value and the mask image of different heterogeneities
are obtained, as shown in Tab.I and Fig.5. The thresholds in
Tab.I respectively represent the average thresholds of each
group of the best heterogeneous mask images in different
wavelength light sources, so as to ensure more accurate
extraction of the feature information of the heterogeneous
region. And as can be seen from Tab.I, the floating range of
while the variation range between different wavelengths is
large. Finally, a total of 9696×14 dimensional information is
obtained for all wavelength images, and each wavelength
includes 2424×14 samples.
2) DATA SET PRODUCTION
All wavelength heterogeneous feature information (9696×14)
is made into a data set, which included a total of 14 data sets.
①Single wavelength test data set, a total of 4 groups. ②Two-
wavelength combined test data set, a total of 6 groups. ③
Three-wavelength combined test data set, a total of 3 groups.
④Four-wavelength combined test data set, a total of 1 group.
To promote the possibility of multi-spectral transmission
images in early breast cancer detection, the wavelength
combination process is sequentially arranged in the order of
blue light, green light, near-infrared light and red light. Data
5
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2019.2930296, IEEE Access

Fulong Liu: Classification of heterogeneity on multi-spectral transmission image based on modulation-demodulation-frame accumulation
and pattern recognition

TABLE I
AVERAGE MASK THRESHOLD OF ALL IMAGES AT DIFFERENT WAVELENGTHS
Image B11 B12 B13 B14 B21 B22 B23 B24 B31 B32 B33 B34

Threshold 0.18 0.16 0.13 0.12 0.14 0.11 0.11 0.10 0.14 0.16 0.14 0.14
Image B41 B42 B43 B44 B51 B52 B53 B54 B61 B62 B63 B64
Threshold 0.15 0.16 0.14 0.12 0.10 0.08 0.09 0.07 0.16 0.14 0.14 0.15
Image G11 G12 G13 G14 G21 G22 G23 G24 G31 G32 G33 G34
Threshold 0.12 0.2 0.07 0.05 0.24 0.1 0.06 0.08 0.18 0.1 0.08 0.07
Image G41 G42 G43 G44 G51 G52 G53 G54 G61 G62 G63 G64
Threshold 0.12 0.38 0.05 0.05 0.32 0.1 0.21 0.2 0.1 0.4 0.07 0.08
Image I11 I12 I13 I14 I21 I22 I23 I24 I31 I32 I33 I34
Threshold 0.42 0.37 0.4 0.47 0.30 0.35 0.31 0.37 0.38 0.43 0.44 0.49
Image I41 I42 I43 I44 I51 I52 I53 I54 I61 I62 I63 I64
Threshold 0.33 0.36 0.35 0.43 0.40 039 0.41 0.42 0.22 0.24 0.25 0.28
Image R11 R12 R13 R14 R21 R22 R23 R24 R31 R32 R33 R34
Threshold 0.36 0.41 0.36 0.45 0.38 0.4 0.36 0.42 0.37 0.39 0.36 0.39
Image R41 R42 R43 R44 R51 R52 R53 R54 R61 R62 R63 R64
Threshold 0.35 0.42 0.35 0.39 0.35 0.38 0.35 0.36 0.2 0.22 0.21 0.19
Note: B11 represents the blue wavelength image of the first heterogeneity in the first group obtained, and so on. G represents the wavelength of green light; I
represents the wavelength of near-infrared light; R represents the wavelength of red light.

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

FIGURE 5. Mask images of all heterogeneities in the near-infrared wavelength image. (a) potato piece 1； (b) potato piece 2； (c) carrot piece 1； (d) carrot
piece 2.

sets of different wavelengths are randomly divided into training
sets and test sets. Since the data set in this paper is small, the
ratio is set to 4:2 according to the traditional division ratio of
the machine learning field. In the solution, the heterogeneities
are divided into 4 types: two pieces of potato and two pieces of
carrot of different sizes and thicknesses.
3) MODEL DEBUGGING
The sample data is normalized during the model debugging
process. In order to find the optimal model, different
combinations of heterogeneous feature information (single
wavelength, two wavelengths, three wavelengths and four
wavelengths) of four wavelengths are input into the network
for training, and each set of data sets is trained 100 times.
According to the classification results of RF and ELM models,
the number of optimal decision trees and neurons in the models
are determined respectively, and the best detection model for
different heterogeneous classifications is obtained. Finally, the
classification accuracy, correlation coefficient R2, root mean
square error (RMSE) and running time of the RF and ELM
models are obtained. The results are shown in Tab.IV.
IV. RESULTS AND ANALYSIS
After the multi-spectral transmission image is processed by the
modulation-demodulation and frame accumulation technique,
the SNR is significantly enhanced and the gray level is
significantly increased, as shown in Fig.6. And it can be seen
from the figure that the gray level of the processed images
increases significantly, which makes the heterogeneous regions
in all wavelength images more prominent. And compared with
the traditional multi-spectral image classification model, the
RF and ELM models established in this paper achieve better
classification accuracy.
(1) Modulation-demodulation and frame accumulation
technique significantly improves the quality of the phantom
image. The results are shown in Tab.II. It is found that by
calculation that the peak signal-to-noise ratio (PSNR) of the
image before and after frame accumulation is positive, which
indicates that the gray level of the image significantly increases
after preprocessing. The image SNR after frame accumulation
has been improved to a certain extent, which will be conducive
to the extraction of regional feature information of
heterogeneity in the image.
(2) Heterogeneous classification results: Based on the
modulation-demodulation and frame accumulation technique

6 VOLUME XX, 2019

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2019.2930296, IEEE Access

Fulong Liu: Classification of heterogeneity on multi-spectral transmission image based on modulation-demodulation-frame accumulation
and pattern recognition

(a1) (a2) (a3) (a4)

(b1) (b2) (b3) (b4)

(c1) (c2) (c3) (c4)

(d1) (d2) (d3) (d4)

FIGURE 6. Gray level comparison chart before and after image preprocessing. (a) blue light image； (b) green light image； (c) near infrared light image；
(d) red light image. a2-d2 are the original phantom images of 4 different wavelengths, and a3-d3 are the pre-processed images. And a1-d1 and a4-d4 are the
gray histogram distribution of a2-d2 and a3-d3 respectively.

TABLE II model for identifying different heterogeneities. As shown in

AVERAGE SNR OF ALL WAVELENGTH IMAGES Fig.7, RF and ELM determine model parameters (number of
Image PSNR/dB SNR1/dB SNR2/dB decision tree and neurons) according to classification accuracy
respectively. It can be seen from the figure that by changing the
Blue wavelength 45.2904 3.4615 3.4708
number of decision trees in RF and the number of neurons in
Green wavelength 39.3314 7.3430 9.3843 ELM respectively, it can be found that the overall classification
Near infrared wavelength 50.4308 4.9463 5.4157 accuracy of the model keeps improving with the increase of the
number of decision trees(neurons). After reaching a certain
Red wavelength 40.4887 5.9977 6.6183
classification accuracy rate, the classification accuracy no
Note: SNR1 represents the SNR of the original image. SNR2 represents the
SNR of the preprocessed image. longer changes, and the optimal number of decision
which has improved the quality of multi-spectral images, the trees(neurons) in the best model is obtained. The results are
data sets of different combined wavelengths are trained in RF shown in Tab.IV. Meanwhile, in order to verify the validity of
and ELM models respectively to obtain the best classification the models we established, we compare its classification

VOLUME XX, 2019 7

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2019.2930296, IEEE Access

Fulong Liu: Classification of heterogeneity on multi-spectral transmission image based on modulation-demodulation-frame accumulation
and pattern recognition

(a) (b)
FIGURE 7. RF and ELM model parameter results image. (a) Determination process of the number of decision trees in the RF model; (b) Process of
determining the number of neurons in the ELM model.
TABLE III
QUALITY FACTOR RESULTS FOR EACH WAVELENGTH COMBINATION VARIABLE
Image a b c d ab ac ad bc bd cd abc abd bcd abcd

Firmness 0.121 0.192 0.040 0.094 0.104 0.163 0.162 0.112 0.164 0.077 0.142 0.190 0.128 0.134
Angular 0.091 0.092 0.180 0.097 0.170 0.125 0.142 0.164 0.115 0.167 0.167 0.119 0.173 0.118
Contrast 0.102 0.126 0.209 0.112 0.181 0.157 0.126 0.170 0.129 0.161 0.138 0.141 0.141 0.122
Correlation 0.143 0.249 0.141 0.089 0.220 0.190 0.181 0.220 0.214 0.114 0.259 0.164 0.207 0.212
Entropy 0.104 0.077 0.113 0.084 0.185 0.167 0.146 0.102 0.102 0.145 0.133 0.138 0.114 0.153
Inverse 0.107 0.125 0.062 0.091 0.135 0.133 0.163 0.081 0.139 0.099 0.117 0.161 0.093 0.147
Smoothness 0.018 0.171 0.176 0.396 0.083 0.183 0.174 0.173 0.266 0.259 0.093 0.175 0.223 0.188
Dissimilarity 0.004 0.014 0.028 0.004 0.014 0.014 0.012 0.020 0.014 0.019 0.007 0.008 0.015 0.007
Consistency 0.001 0.025 0.029 0.015 0.019 0.016 0.018 0.030 0.029 0.019 0.020 0.016 0.022 0.009
Center 0.001 0.033 0.028 0.021 0.017 0.011 0.018 0.046 0.032 0.026 0.032 0.019 0.038 0.019
Area 0.001 0.035 0.033 0.013 0.015 0.009 0.012 0.037 0.025 0.018 0.022 0.019 0.028 0.016
Perimeter 0.008 0.002 0.004 0.002 0.002 0.002 0.002 0.006 0.007 0.003 0.005 0.004 0.009 0.005
Long 0.006 0.034 0.017 0.001 0.013 0.012 0.010 0.021 0.023 0.006 0.021 0.019 0.032 0.017
Short 0.003 0.038 0.010 0.002 0.016 0.015 0.016 0.026 0.030 0.005 0.029 0.018 0.019 0.019
Note: ①The green mark indicates the factor with the highest weight of the variables in the different wavelength combinations. ②The red mark indicates the highest
weight of the same factor in different wavelength combinations.

accuracy with the traditional multi-spectral image
classification models, and the results are shown in Tab.Ⅴ. In
addition, based on the Gini index in RF, the importance of each
characteristic variable in the model prediction process is also
obtained, as shown in Tab.III. The specific analysis process of
the importance of characteristic variables based on Gini index
is as follows:
(1) The importance of feature x i on node n , that is, the
Gini index change before and after the data on node n
is divided into its left and right sub-nodes n l and n r
is shown in equation (2):
IM PinGini  I G ( n )  I G ( nl )  I G ( n r )
(2)
(2) If the node set of feature x i as node segmentation
attribute in the k -th decision tree is N , the importance
of feature in this decision tree can be obtained from
equation (3):
k 
IM Pi G_ ini  n N IM PinG ini
(3)
(3) If there are K trees in RF, the importance of feature x i
in the whole RF can be calculated by equation (4):
1 K
IMPi Gini   IMPi Gini
K k 1
_k
(4)
Where: IM PinG in i represents the average change of Gini index
of the i -th feature on all RF decision tree nodes; I G ( n )
represents the exponential change of node n ; I G (nl ) and
I G ( nr ) represent respectively the change of Gini index before
and after distributing the data on node n to its left and right
sub-nodes nl and n r ; IM Pi G_ ink i represents the importance
of the i -th characteristic variables in the k -th decision tree
(the set of nodes where the node partitioning attribute is N );
IM Pi G in i represents the importance of the i -th characteristic
variable in the whole RF.
As shown in Tab.III, the importance of all characteristic
variables in the model establishment process is obtained based
on Gini index. It can be seen from table that in the model

8 VOLUME XX, 2019

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2019.2930296, IEEE Access

Fulong Liu: Classification of heterogeneity on multi-spectral transmission image based on modulation-demodulation-frame accumulation
and pattern recognition

TABLE IV
RF AND ELM MODEL OPERATION RESULTS
Model RF ELM
2 2
Test Training% Test% R RMSE Time/s DT Var Training% test% R RMSE Time/s Neurons Var
a 92.08 78.34 0.34 1.03 760.51 3 2.73 88.37 85.77 0.54 0.85 146.01 4 2.39
b 99.94 96.91 0.98 0.18 1437.44 3 1.05 92.33 91.09 0.92 0.29 147.27 45 2.04
c 99.01 89.60 0.87 0.41 936.12 3 2.21 93.81 92.57 0.88 0.39 145.02 27 2.37
d 100 93.69 0.95 0.25 1302.29 3 1.54 89.29 88.24 0.91 0.34 145.32 28 2.48
ab 100 100 1 0 1172.74 5 0 100 99.38 0.98 0.15 145.99 28 0.35
ac 100 100 1 0 1395.82 5 0 100 100 1 0 149.04 35 0
ad 100 100 1 0 1085.09 5 0 100 100 1 0 148.57 36 0
bc 100 99.01 0.99 0.099 1491.12 5 0.59 96.53 96.04 0.97 0.20 151.02 21 1.14
bd 100 100 1 0 1656.84 5 0 97.40 97.65 0.98 0.15 147.20 25 0.82
cd 100 97.65 0.98 0.15 1798.54 5 0.76 97.34 94.80 0.95 0.24 151.46 23 1.28
abc 100 100 1 0 1729.82 6 0 100 100 1 0 154.76 19 0
abd 100 100 1 0 1544.67 6 0 100 100 1 0 150.79 19 0
bcd 100 99.88 0.98 0.13 1994.85 6 0.53 98.02 97.65 0.98 0.15 154.47 27 0.68
abcd 100 100 1 0 1501.31 6 0 100 100 1 0 158.87 23 0
Note: ‘a’ represents the blue wavelength. ‘b’ represents the green wavelength. ‘c’ represents the near-infrared wavelength. And ‘d’ represents the red wavelength.
DT=Decision trees. DT represents the number of optimal decision trees in each model. Neurons represents the number of optimal neurons in each model. Var=
variance.
TABLE Ⅴ
AVERAGE ACCURACY OF DIFFERENT CLASSIFICATION MODELS
Wavelength\Model SVM LDA LRR SR MLR RF1 ELM1 RF ELM
Sigle 89.51 82.04 88.86 84.60 89.04 57.79 55.34 89.64 89.42
Two 90.89 82.74 83.88 83.25 97.14 62.65 60.05 99.44 97.98
Three 98.78 92.15 71.72 67.74 99.12 63.08 64.4 99.96 99.22
Four 99.90 93.28 84.28 81.70 93.93 73.84 76.08 100 100
Note: RF1 and ELM1 respectively represent the classification accuracy of heterogeneity in unpreprocessed multi-spectral transmission images. RF and ELM
respectively represent the classification accuracy of heterogeneity in preprocessed multi-spectral transmission images.

building process, the gray-scale correlation, smoothness,
contrast and firmness of the variables extracted from the multi-
spectral transmission images account for a relatively high
proportion. And the same variable factor distribution with a
higher specific gravity in different wavelength combinations is
relatively dispersed, but mainly concentrated in single
wavelength and two wavelength combinations. This indicates
that as the wavelength combination increases, the proportion of
each characteristic variable tends to be equalized during the
model establishment process, enriching the information degree
of the heterogeneities.
The average results of classification after multiple training
of the model are shown in Tab.IV. Under the experimental data,
the classification results of the RF and ELM models are
compared by the accuracy of the classification, R2, RMSE and
the running time. Among them, R2 and RMSE are calculated
by formulas (5) and (6). It can be seen from Tab.IV that as the
the establishment of the optimal ELM model, the number of
neurons in ELM is relatively disorderly distributed. And
compared with single-network model, multi-network model
takes more time to build. Therefore, the following conclusions
can be drawn: ①The overall classification of RF and ELM is
ideal. ②The classification effect of the combined wavelength
is better than the single-wavelength classification effect, and
the four-wavelength combination classification effect is
optimal. ③With the increase of wavelength, the classification
accuracy of the models (RF and ELM) is gradually improved.
④ In the classification effect of different wavelength
combinations, RF is superior to ELM. ⑤ In the model
classification running time, ELM is shorter than RF, and the
difference of time amount is large. ⑥Under the premise of
small difference in overall classification accuracy, ELM's
operation efficiency is significantly higher than RF.
m

combined wavelength information increases, the classification  ( y  y )

i i
2

accuracy of the models (RF and ELM) are also increasing. At R  1

2 i 1
m
(5)
the same time, with the increasing of combined wavelength ( y  y ) i i
2

information, the model's fitting correlation R2 is also improving, i 1

and RMSE and variance are decreasing, which indicates that 1 m


the model's stability is improving and the over-fitting
RM SE 
m
 (y
i 1
i yi ) 2 (6)

phenomenon of the model is reasonably suppressed. Moreover,
as can be seen from Tab.IV, in the process of obtaining the best
RF model, the number of classification decision trees gradually
increases with the increase of wavelength combination. But in
Where m represents the number of test samples;  yi
represents the predicted value of the test sample; yi represents
the actual value of the test sample; y i represents the actual

VOLUME XX, 2019 9

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2019.2930296, IEEE Access
Fulong Liu: Classification of heterogeneity on multi-spectral transmission image based on modulation-demodulation-frame accumulation
average value of the test sample.
The average classification accuracy of different models is
shown in Tab.Ⅴ. It can be seen from the table that compared
with the RF and ELM models established by multi-spectral
transmission images before preprocessing, the classification
accuracy of all wavelength combination RF and ELM models
established by multi-spectral transmission images after
preprocessing is significantly improved, which indicates the
effectiveness of the proposed method to enhance image quality.
And as can also be seen from the table that in the models of
single wavelength and four wavelength combination, the
traditional SVM classification effect is the best, reaching 89.51
and 99.90, respectively. In the models of the two and three
wavelength combination, the traditional MLR classification is
the best, reaching 97.14 and 99.12, respectively. However,
compared with the combination of different wavelengths, the
RF and ELM models established after the preprocessed of
multi-spectral transmission images in this paper have better
classification effect, and the RF and ELM models have better
overall classification effect.
Analysis of experimental results: The image preprocessed by
the modulation-demodulation and frame accumulation
technique, it is found that the SNR of the image is improved to
a certain extent, which makes the heterogeneities in the image
more prominent and the extracted feature information more
perfect. With the same error rate and iteration times, RF's
classification accuracy is slightly higher than ELM, while ELM
is slightly superior in classification speed. Therefore, ELM
operates slightly more efficiently than RF. And compared with
the accuracy of traditional multi-spectral image classification
models, the RF and ELM we established have better
classification effect.
V. CONCLUSION
In this paper, combined with the characteristics of breast tissue,
multi-spectral transmission image acquisition experiment is
designed. The image quality is improved by modulation-
demodulation and frame accumulation technique, and the
heterogeneous classification is studied by pattern recognition
(RF, ELM) and traditional multi-spectral image classification
models (SVM, LDA, LRR, SR and MLR). Firstly, in the
process of acquiring multi-spectral transmission images,
sinusoidal signals with frequency of 4HZ are used as carrier
signals to enhance the information degree of the image. Then,
in image preprocessing, the PSNR and SNR of the image are
improved through the modulation-demodulation and frame
accumulation technology. The result shows that the PSNR of
the image before and after frame accumulation is positive,
which indicates that the gray level of the image significantly
increases after preprocessing. And the image SNR after frame
accumulation has also been improved to a certain extent, which
will obtain more abundant feature information of the image
heterogeneous region. Finally, in the aspect of heterogeneous
classification, the different wavelength combination data sets
of phantom feature information are trained as the training
sample, and the heterogeneous classification accuracy of
different models is compared. The results show that compared
with the classification accuracy of the traditional multi-spectral
image classification models, RF and ELM models have better
classification effect. Among them, the RF and ELM models
10
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
and pattern recognition
established in the data set of four-wavelength combination
have the best classification effect, and the classification
accuracy is up to 100%, secondly, it is the three-wavelength
combined model. And the single-wavelength model has the
worst classification effect. And, in terms of model operation
efficiency, ELM is significantly higher than RF. In summary,
the technique of modulation-demodulation and frame
accumulation has improved the gray resolution of images. And
compared with the classification accuracy of the traditional
multi-spectral image classification models, the RF and ELM
models established in this paper have better classification effect.
In addition, potatoes and carrots are selected as heterogeneities
to simulate breast cancer according to the characteristics of
strong transmissibility and tomographic distribution of breast
tissue, which may be limited by other unknown conditions, but
provides a good idea for the detection of heterogeneity in breast
tissue. Through further research, the improved method can be
adapted to more complex situations and even clinical
applications. It is expected that this method can promote the
clinical application of multi-spectral transmission imaging in
early screening of breast tumors in the future.
ACKNOWLEDGMENT
We thank the State Key Laboratory of Precision Measuring
Technology and Instruments for the use of their equipment.
FUNDING
This research did not receive any specific grant from funding
agencies in the public, commercial, or not-for-profit sectors.
CONFLICT OF INTEREST
The authors declare that they have no conflict of interest.
REFERENCES
[1] Joëlle A.C. Desreux, “Breast cancer screening in young women,”
European Journal of Obstetrics & Gynecology and Reproductive Biology,
vol. 230, pp. 208-211, 2018.
[2] Sharon W.W. Chan et al., “Surgical options for Chinese patients with early
invasive breast cancer: Data from the Hong Kong Breast Cancer Registry,”
Asian Journal of Surgery, vol. 40, no. 6, pp. 444-452, 2017.
[3] Zeng Jie Ye et al., “A biopsychosocial model of resilience for breast
cancer: A preliminary study in mainland China,” European Journal of
Oncology Nursing, vol. 36, pp. 95-102, 2018.
[4] C. Sosa et al., “Compact energy dispersive X-ray microdiffractometer for
diagnosis of neoplastic tissues,” Radiation Physics and Chemistry, vol.
137, pp. 125-129, 2017.
[5] Damien Barbes et al., “Material-specific imaging system using energy-
dispersive X-ray diffraction and spatially resolved CdZnTe detectors with
potential application in breast imaging,” Nuclear Instruments and
Methods in Physics Research Section A: Accelerators, Spectrometers,
Detectors and Associated Equipment, vol. 848, pp. 91-98, 2017.
[6] Daisuke Shimao et al., “Imaging with ultra-small-angle X-ray scattering
using a Laue-case analyzer and its application to human breast tumors,”
Physica Medica, vol. 44, pp. 236-242, 2017.
[7] Min Xian et al., “Automatic breast ultrasound image segmentation: A
survey,” Pattern Recognition, vol. 79, pp. 340-355, 2018.
[8] Huay-Ben Pan, “The Role of Breast Ultrasound in Early Cancer
Detection,” Journal of Medical Ultrasound, vol. 24, no. 4, pp. 138-141,
2016.
[9] Rongrong Guo et al., “Ultrasound Imaging Technologies for Breast
Cancer Detection and Management: A Review,” Ultrasound in Medicine
& Biology, vol. 44, no. 1, pp. 37-70, 2018.
[10] Shuang-Qing Chen et al., “Application of Abbreviated Protocol of
Magnetic Resonance Imaging for Breast Cancer Screening in Dense
Breast Tissue,” Academic Radiology, vol. 24, no. 3, pp. 316-320, 2017.
[11] Albert Gubern-Mérida et al., “Automated detection of breast cancer in
false-negative screening MRI studies from women at increased risk,”
VOLUME XX, 2019
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2019.2930296, IEEE Access
Fulong Liu: Classification of heterogeneity on multi-spectral transmission image based on modulation-demodulation-frame accumulation
European Journal of Radiology, vol. 85, no. 2, pp. 472-479, 2016.
[12] Seyedamir Tavakoli Taba et al., “Toward Improving Breast Cancer
Imaging: Radiological Assessment of Propagation-Based Phase-Contrast
CT Technology,” Academic Radiology, 2018.
[13] Ying Yang et al., “Near-infrared light triggered liposomes combining
photodynamic and chemotherapy for synergistic breast tumor therapy,”
Colloids and Surfaces B: Biointerfaces, vol. 173, pp. 564-570, 2019.
[14] R.G. Pleijhuis et al., “Near-infrared fluorescence (NIRF) imaging in
breast-conserving surgery: Assessing intraoperative techniques in tissue-
simulating breast phantoms,” European Journal of Surgical Oncology
(EJSO), vol. 37, no. 1, pp. 32-39, 2011.
[15] Meihua Zhang et al., “Near-infrared photothermal therapy using EGFR-
targeted gold nanoparticles increases autophagic cell death in breast
cancer,” Journal of Photochemistry and Photobiology B: Biology, vol.
170, pp. 58-64, 2017.
[16] Xue Yang et al., “Assessment of spatial information for hyperspectral
imaging of lesion,” International Society for Optics and Photonics, 2016.
[17] Francisco Peñaranda et al., “Multimodal registration of optical
microscopic and infrared spectroscopic images from different tissue
sections: An application to colon cancer,” Digital Signal Processing, vol.
68, pp. 1-15, 2017.
[18] Jonathan W. Uhr et al., “Molecular profiling of individual tumor cells by
hyperspectral microscopic imaging,” Translational Research, vol. 159, no.
5, pp. 366-375, 2012.
[19] P. Musto et al., “Hyperspectral Raman imaging of human prostatic cells:
An attempt to differentiate normal and malignant cell lines by univariate
and multivariate data analysis,” Spectrochimica Acta Part A: Molecular
and Biomolecular Spectroscopy, vol. 173, pp. 476-488, 2017.
[20] Qingli Li et al., “Methyl green and nitrotetrazolium blue chloride co-
expression in colon tissue: A hyperspectral microscopic imaging analysis,”
Optics & Laser Technology, vol. 64, pp. 337-342, 2014.
[21] S. Dontu et al., “Combined spectral-domain optical coherence
tomography and hyperspectral imaging applied for tissue analysis:
Preliminary results,” Applied Surface Science, vol. 417, pp. 119-123,
2017.
[22] Li G et al., “Employment of frame accumulation and shaped function for
upgrading low-light-level image detection sensitivity,” Optics Letters,
vol.37, no. 8, pp. 1361-1363, 2012.
[23] Yang X et al., “Effect on measurement accuracy of transillumination using
sawtooth-shaped-function optical signal,” Review of Scientific
Instruments, vol. 87, no. 11, pp. 115106, 2016.
[24] Hu YJ et al., “Optimum method of image acquisition using sawtooth-
shaped-function optical signal to improve grey-scale resolution,” Journal
of Modern Optics, vol. 63, pp. 1539-1543, 2016.
[25] G. Camps-Valls et al., “Spatio-spectral remote sensing image
classification with graph kernels,” IEEE Geosci. Remote Sens. Lett, vol.
7, no. 4, pp. 741–745, Oct. 2010.
[26] X. Guo et al., “Support tensor machines for classification of hyperspectral
remote sensing imagery,” IEEE Trans. Geosci. Remote Sens, vol. 54, no.
6, pp. 3248–3264, Jun. 2016.
[27] H. R. Shahdoosti et al., “Spectral–spatial feature extraction using
VOLUME XX, 2019
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
and pattern recognition
orthogonal linear discriminant analysis for classification of hyper-spectral
data,” Eur. J. Remote Sens, vol. 50, no. 1, pp. 111–124, 2017.
[28] H. Wan et al., “A new linear discriminant analysis method to address the
over-reducing problem,” in Proc. Int.Conf. Pattern Recognit. Mach. Intell,
pp. 65–72, 2015.
[29] Y. Li et al., “Learning low-rank representations with classwise block-
diagonal structure for robust face recognition,” in Proc. AAAI, pp. 2810–
2816, 2014.
[30] L. Li et al., “Learning low-rank and discriminative dictionary for image
classification,” Image Vis. Comput, vol. 32, no. 10, pp. 814–823, 2014.
[31] G. Liu et al., “Robust recovery of subspace structures by low-rank
representation,” IEEE Trans. Pattern Anal. Mach. Intell, vol. 35, no. 1, pp.
171–184, Jan. 2013.
[32] C. Li et al., “Hyperspectral image classification with robust sparse
representation,” IEEE Trans. Geosci. Remote Sens, vol, 13, no. 5, pp.
641–645, Mar. 2016.
[33] J. Li et al., “Efficient superpixel-level multi-task joint sparse
representation for hyperspectral image classification,” IEEE Trans.
Geosci. Remote Sens, vol. 53, no. 10, pp. 5338–5351, Oct. 2015.
[34] L. Xu et al., “Bayesian classification of hyperspectral imagery based on
probabilistic sparse representation and Markov random field,” IEEE
Geosci. Remote Sens. Lett, vol. 11, no. 4, pp. 823–827, Apr. 2014.
[35] J. Li et al., “Semisupervised hyperspectral image classification using soft
sparse multinomial logistic regression,” IEEE Geosci. Remote Sens. Lett,
vol. 10, no. 2, pp. 318–322, Mar. 2013.
[36] W. Gao et al., “Ideal kernel-based multiple kernel learning for spectral-
spatial classification of hyperspectral image,” IEEE Geosci. Remote Sens.
Lett, vol. 14, no. 7, pp. 1051–1055, Jul. 2017.
[37] Q. Wang et al., "Locality Adaptive Discriminant Analysis for Spectral–
Spatial Classification of Hyperspectral Images," in IEEE Geoscience and
Remote Sensing Letters, vol. 14, no. 11, pp. 2077-2081, Nov. 2017.
[38] Q. Wang et al., "Locality and Structure Regularized Low Rank
Representation for Hyperspectral Image Classification," in IEEE
Transactions on Geoscience and Remote Sensing, vol. 57, no. 2, pp. 911-
923, Feb. 2019.
[39] Q. Gao et al., "Hyperspectral Image Classification Using Joint Sparse
Model and Discontinuity Preserving Relaxation," in IEEE Geoscience
and Remote Sensing Letters, vol. 15, no. 1, pp. 78-82, Jan. 2018.
[40] BREIMAN L, “Random forest,” Machine Learning, vol. 45, no. 1, pp. 5-
32, 2001.
[41] Huang G B et al., “Extreme learning machine: theory and applications,”
Neurocomputing, vol. 70, no. 1-3, pp. 489-501, 2006.
[42] Gang Li et al., “Optimum method of applying and removing a shaped-
function signal for low-light-level image detection,” Appl. Opt, vol. 52,
pp. 7934-7941, 2013.
[43] ALTMANN A et al., “Permutation importance: a corrected feature
importance measure,” Bioinformatics, vol. 26, no. 10, pp. 1340-1347,
2010.
[44] Hamed Vavadi et al., “Compact ultrasound-guided diffuse optical
tomography system for breast cancer imaging,” JOURNAL OF
BIOMEDICAL OPTICS, vol. 24, no. 2, 2019.
11