A Novel End-To-End 1D-ResCNN Model To Remove Artifact From EEG Signals
A Novel End-To-End 1D-ResCNN Model To Remove Artifact From EEG Signals
A Novel End-To-End 1D-ResCNN Model To Remove Artifact From EEG Signals
Neurocomputing
journal homepage: www.elsevier.com/locate/neucom
a r t i c l e i n f o a b s t r a c t
Article history: Electroencephalography (EEG) signals are an important tool in the field of clinical medicine, brain re-
Received 18 February 2019 search and the study of neurological diseases. EEG is very susceptible to a variety of physiological sig-
Revised 5 November 2019
nals, which brings great difficulties to the research and analysis of EEG signals. Therefore, removing noise
Accepted 8 April 2020
from EEG signals is a key prerequisite for analyzing EEG signals. In this paper, a one-dimensional residual
Available online 23 April 2020
Convolutional Neural Networks (1D-ResCNN) model for raw waveform-based EEG denoising is proposed
Communicated by Dr. Nianyin Zeng to solve the above problem. An end-to-end (i.e. waveform in and waveform out) manner is used to map
a noisy EEG signal to a clean EEG signal. In the training stage, an objective function is often adopted to
Keywords:
Electroencephalogram (EEG) optimize the model parameters and in the test stage, the trained 1D-ResCNN model is used as a filter
Artifacts removal to automatically remove noise from the contaminated EEG signal. The proposed model is evaluated on
Deep learning the EEG signal from the CHB-MIT Scalp EEG Database, and the added noise signals are obtained from the
End-to-end database. We compared the proposed model with the independent of the composite analysis (ICA), the
One-dimensional residual convolutional fast independent composite analysis (FICA),Recursive least squares(RLS) filter,Wavelet Transform (WT) and
neural networks model (1D-ResCNN) Deep neural network(DNN) models. Experimental Results show that the proposed model can yield cleaner
waveforms and achieve significant improvement in SNR and RMSE.Meanwhile, the proposed model can
also preserve the nonlinear characteristics of EEG signals.
© 2020 Published by Elsevier B.V.
https://doi.org/10.1016/j.neucom.2020.04.029
0925-2312/© 2020 Published by Elsevier B.V.
W. Sun, Y. Su and X. Wu et al. / Neurocomputing 404 (2020) 108–121 109
Nonlinear adaptive filters further include Volterra filters and neural structure in the input EEG signal, with weight sharing and defor-
network based adaptive filters [12–17]. Nonlinear adaptive filters mation robustness.Although they have inherent advantages in EEG
have stronger signal processing capabilities and complex calcula- analysis, the CNN model has not received sufficient attention in
tion [17,18]. Linear adaptive filters are too sensitive and unsTable EEG noise reduction.
to adjust the parameters. Most importantly, artifacts overlap most EEG signals are usually long one-dimensional complex sig-
of the clean EEG signals [19,20]. Therefore, filters may eliminate nals.Due to the time-varying and complexity of EEG signals, more
useful EEG signals during artifact deletion. complex nonlinear features should be extracted for EEG artifact
Theempirical mode decomposition(EMD) proposed by Huang removal.Therefore,in order to overcome the shortcomings of the
[21] decomposes the input signals into multiple empirical modes above traditional methods,and considering the nonlinear character-
according to the inherent mode (IMFs) function, which is benefi- istics of EEG time-varying signals and the advantages of CNN fea-
cial to the analysis of multi-component signals.EMD is an empirical ture extraction, this paper proposes a new one-dimensional resid-
and data-driven method for dealing with non-stationary, nonlinear, ual convolutional neural network model (1D-ResCNN) based on
stochastic processes, so it is well suited for EEG signal analysis and multi-scale kernel to remove noise from the EEG signal.The pro-
processing.However,EMD is computationally complex and may not posed model can automatically learn the nonlinear and discrimi-
be suiTable for online applications [22,23]. native deep features of the noisy EEG data and true EEG data.Then,
Blind source separation (BSS) is one of the most popular artifact these features are used to distinguish them and automatically re-
removal methods [24–27]. Independent Component Analysis (ICA) construct to obtain a clean EEG signal.
is a multi-dimensional signal processing method developed from The main contributions of this paper are as follows:
BBS, which can separate the ideal signal and noise included in the
(1) In the absence of sufficient prior knowledge, a new 1D-
EEG signal as independent components to achieve EEG signal de-
ResCNN based brain signal denoising model is proposed,
noising [28,29]. Many BSS algorithms require human intervention
which is the first application of CNN in EEG denoising;
to identify artifact components. This makes it subjective and time
(2) The proposed model operates directly on the raw EEG sig-
consuming [28,29].
nal without pre-processing or manual feature extraction.Mo-
The wavelet transform(WT) maps the signal to the wavelet do-
reover, the EEG nonlinear characteristics and the details of
main. According to the wavelet coefficients of signal and noise,
the waveform are also kept;
they have different properties and mechanisms at different scales,
(3) Combining the residual blocks of different scales, the model
eliminating the wavelet coefficients generated by noise and maxi-
obtains more abundant features and increases the nonlinear
mally retaining the coefficients of real signals [30–32].
expression ability of the convolutional neural network.
EEG signal is a complex chaotic signal with nonlinear character-
istics, and the preservation of nonlinear features is of great signif- The structure of this paper is as follows. Section 2 describes the
icance for EEG analysis and classification. Therefore, an EEG noise structure of the proposed 1D-ResCNN; Section 3 gives the experi-
reduction method is needed to make the denoised EEG signal still mental settings and datasets; Section 4 gives the experimental re-
maintains non-linear characteristics. On the other hand, with the sults; Finally, Section 5 shows the conclusion.
advancement of technology, the collection of EEG data has be-
come more convenient, providing a strong support for the imple- 2. Proposed 1D-ResCNN model
mentation of deep learning. In recent years, deep learning tech-
niques have been able to learn high-level and hierarchical repre- In this section, a deep network structure including residual
sentations directly in massive raw data, thus achieving a series of Convolutional Neural Network is designed for long duration seg-
breakthroughs in signal processing. Specifically, Xu et al. [33] ap- ment denoising of EEG signals.The designed denoising system has
plied Deep Neural Networks(DNN) to speech enhancement, and a complete end-to-end structure that does not require feature ex-
proposed a DNN-based minimum mean square error regression fit- traction of the signal at any stage.The input of the network struc-
ting speech enhancement algorithm based on logarithmic power ture are long segments of original EEG signal.The reconstruction of
spectrum of the complex relationship between noisy speech and the noisy signal has been provided at the network output.
clean speech. Xu et al. [33] and Rodrigues and Couto [34] intro-
duces an Restricted Boltzmann Machine(RBM)-based ECG denois- 2.1. Application of CNN in EEG artifact removal
ing method, and [35] proposes an improved denoising automatic
encoder (DAE) improved by wavelet transform (WT) for ECG de- The brain is a complex nonlinear system.EEG is a complex elec-
noising.Yang [36] introduced a Deep Learning Networks(DLN) EEG trophysiological signal that usually has nonlinearities,diversity and
denoising method, which subtly utilizes the structural features of uncertainty that are difficult to process by linear methods.There-
deep learning and powerful learning capabilities to improve EEG fore,nonlinear denoising for EEG signal is imperative.
denoising due to EOG artifact. Since the EEG signal is a one-dimensional discrete time se-
Convolutional Neural Networks(CNN) is a subset of deep learn- ries,we propose a 1D-ResCNN model that extracts the EEG signal
ing, which has attracted a lot of attention in recent years [37,38]. characteristics using 1D convolutional layer. Fig. 1 shows the over-
It has been used in other fields for raw continuous signals, and all proposed 1D-ResCNN structure for EEG signal waveform denois-
it works well, starting with image applications, followed by many ing, due to the small amount of EEG data, the data is enhanced
other fields, such as natural language processing, speech processing with a fixed-size sliding window [43]. The data is enhanced by
[39]. In contrast, CNN networks have not yet been used in remove dividing the given full length EEG signal into sub-signals using a
artifacts from EEG signals. However, CNN networks have recently fixed-size sliding window.We use a window size of 400 with a
found applications in studies focusing on other areas of EEG time stride of 380 (%95 of 400); each record is divided into multiple
series analysis. In works [40–42] CNN is implemented on EEG sig- equal EEG sub-signals, and the last sample is discarded. The sub-
nals, and the effectiveness of CNN algorithm in signal analysis is signal is used as the input to the 1D-ResCNN model, 1D-ResCNN
studied. In general, CNN has several natural advantages.First, CNN model uses the M convolutional layers, each of which has a 1 × N
is superior to traditional methods not only in accuracy but also in convolution kernel to learn the noisy EEG features,where Filter-
speed.Second,CNN can automatically extract and learn the best fea- M − N represents the Nth filter in layer M. Each filter is con-
tures from the original signal to achieve adaptive design.The most volved with all generated waveforms from the previous layer and
important point is that CNN is good at mining the spatio-temporal produces a further filtered waveform EEG.Finally, these features
110 W. Sun, Y. Su and X. Wu et al. / Neurocomputing 404 (2020) 108–121
Fig. 1. The architecture of the training phase of the EEG denoising model based on the one-dimensional residual convolutional neural network (1D-ResCNN). First, a fixed-
size sliding window is used to divide the given full-length EEG signal into small segments to enhance the data. The colored frame overlap represents the signal overlap
rate, and each small signal is used as a separate example of learning the 1D-ResCNN model. The 1D-ResCNN model uses the M convolutional layers, each of which has a
1 × N convolution kernel to learn the noisy EEG features. Finally, these features that can distinguish EEG signals and noise are automatically reconstructed by minimizing
the objective function(MSE) to get the clean EEG signal.
Fig. 2. 1D-ResCNN network structure. Resblock1, Resblock2, and Resblock3 represent the residual blocks of the convolution kernels of 3∗ 1, 5∗ 1, and 7∗ 1, respectively, and the
three residual blocks are connected in series twice, and finally the outputs of the three multi-scale residual blocks are merged. And through a convolution layer, the final
reconstruction signal is obtained.
that can distinguish EEG signals and noise are automatically re- main layers: batch normalized layers(BN), convolutional layers, and
constructed by minimizing the objective function(MSE) to get the Inception-ResNet.
clean EEG signal.This deep network is end-to-end (noisy EEG signal BN layer: During training, changes in network parameters can
waveform and clean EEG signal waveform) mode,providing auto- cause changes in the distribution of input data later, which greatly
matic noise reduction and inputting EEG waveform segments with- reduces the generalization ability and training speed of the net-
out any manual feature extraction or post-processing. work, and makes saturation nonlinearity more difficult. This phe-
nomenon is called Internal Covariate Shift, Ioffe and Szegedy
[44] proposed batch normalization (BN) to solve this problem. For
2.2. 1D-ResCNN network architecture each batch, the activation output of the layer is normalized using
a batch normalization layer, which helps to avoid special initializa-
The structure of the 1D-ResCNN module in this paper is divided tion of parameters, speed up training, optimize results, and provide
into three sub-modules. The first is the data enhancement module, faster convergence. In order to optimize the training process of 1D-
which uses the sliding window method to divide its EEG signal ResCNN, our proposed model performs BN after the convolutional
data into sub-signals [43] using a fixed-size overlapping window. layer.
Then, its 1D-ResCNN extracts features by convolving each sub- 1D-Conv +ReLU layer: In view of the time-varying characteris-
signal. The convolution operation of our proposed method con- tics of EEG, 1D-ResCNN model is constructed considering the ob-
sists of two convolutional layers and an Inception module, which servation of the local features of the convolution kernel. The in-
has both vertical extension and horizontal expansion. The tradi- put feature in 1D-ResCNN model is a one-dimensional vector, the
tional CNN only has a vertical extension. Finally, the features ex- one-dimensional convolution kernel can be regarded as a sliding
tracted by the convolution operation are convolved by the network window on a time series, and the short-term features between
to generate a clean EEG signal. The overall model of 1D-ResCNN is sequences are extracted to achieve the effect of processing time-
shown in Fig. 2. In this study, the 1D-ResCNN is made up of 3 con- varying features. In the first layer of our proposed 1D-ResCNN
volutional layers,3 batch normalized layers (BN) and 1 Inception- model, a 1D convolution is performed on the input EEG raw signal,
ResNet.In the rest of this section, we will briefly introduce the the number of kernels is 16, and receptive field for each kernel is
W. Sun, Y. Su and X. Wu et al. / Neurocomputing 404 (2020) 108–121 111
Table 1
Detailed parameters used for all the layers of
proposed CNN model.
Input – (None,400,1)
Conv1D 1 × 5,32 (None,400,32)
BN – (None,400,32)
ResBlock1 1 × 3 (None,400,32)
ResBlock1 1 × 3 (None,400,32)
add – –
ResBlock2 1 × 5 (None,400,32)
ResBlock2 1 × 3 (None,400,32)
ResBlock3 1 × 5 (None,400,32)
ResBlock3 1 × 3 (None,400,32)
merge – (None,400,96)
Conv1D 1 × 1,64 (None,400,1)
BN – (None,400,1)
Dense 1 (None,400,1)
Fig. 4. The flow chart of the 1D-ResCNN denoising model mainly includes training and testing. In the training phase, the network automatically learns the nonlinear
and discriminative deep features in the noisy EEG data collected. Finally, the network utilizes These features that distinguish the EEG signal and noise are automatically
reconstructed to obtain a clean EEG signal. In the test phase, the test focused noisy EEG signal is input into the trained 1D-ResCNN network to obtain a clean EEG signal.
Fig. 5. A typical contaminated EEG signal for 8s.The signal is divided into 10 windows in the same time period, and the excess is removed. Noise mainly exists in window
1, window 3, window 5, window 7 and window 9, while the EEG segments in the remaining windows can be considered to contain no noise.
Fig. 7. Comparison of spectrograms: (a) the noisy EEG signal of the ECG noise at
0 dB, and the speech enhanced by (b) 1D-ResCNN, (c) DNN, (d) WT, (e) FICA, (f)
ICA, (g) RLS filter and (h) the clean EEG signal.
is divided into 10 windows in the same time period, and the ex-
cess is removed. Fig. 5 shows that noise mainly exists in window 1,
window 3, window 5, window 7 and window 9, while the EEG seg-
ments in the remaining windows can be considered to contain no
noise. Calculate the kurtosis of the EEG segment in each window
and obtain the kurtosis value kw (w indicates the serial number of
the window). Then based on a certain amount of observation and
empirical analysis, set the threshold K to pick up the EEG segment.
If the EEG segment kurtosis kw exceeds the threshold K, this seg-
ment is eliminated. The remaining EEG segments are considered
uncontaminated EEG segments. In this way, we can get enough
clean EEG.
RMS(λnoise )
SNR = (8)
RMS(E E Gc )
Where RMS is defined as the root mean square energy of the sig-
nal.
These synthetically contaminated EEG signals are further used
to correct the evaluation of the algorithm and change the intensity
of the contamination to verify the efficiency of the algorithm.
Fig. 6. An example of an experimental result waveform for eliminating noise from 3.2.1. Unknown noise test dataset
an EEG signal containing ECG noise of SNR = 0 dB. From top to bottom, the re-
In order to test the generalization ability of the EEG noise re-
sults are denoised by ICA, FICA, RLS filter, WT, DNN and the method proposed in
this paper. The dotted line represents the noisy EEG waveform, and the solid line duction model, an unknown noise test set is constructed. The two
represents the EEG waveform after using different denoising methods. types of noise in the unknown noise test set are unknown noises
completely different from the training set noise, namely Gaussian
White Noise (WGN), and electrode motion(EM). In order to make
3.1. Intercepting clean EEG training samples the experimental conditions more realistic, the noise type and SNR
level of the training and test sets do not match. The EEG signal of
In this paper, we use the EEG samples without noise to train the test set and the randomly intercepted segment of the 2 types
the 1D-ResCNN model. Inspired by [36], We used a clean EEG sig- of noise are mixed according to the global SNR of −10 dB, 0 dB
nal based on kurtosis interception. Fig. 5 is a typical contaminated and 10 dB, respectively, to obtain a test set of noisy brain electric-
EEG signal within 8 s (with sampling frequency 256 Hz).The signal ity with unknown noise.
114 W. Sun, Y. Su and X. Wu et al. / Neurocomputing 404 (2020) 108–121
Fig. 9. Comparison of spectrograms: (a) the noisy EEG signal of the EMG noise at
0 dB, and the speech enhanced by (b) 1D-ResCNN, (c) DNN, (d) WT, (e) FICA, (f)
ICA, (g) RLS filter and (h) the clean EEG signal.
where x(i) is the original EEG signal, then x(ˆi ) represents the de-
noised EEG signal.
The RMSE defines the energy of the error signal during noise
reduction. A lower value for RMSE means better estimation of the
original signal and better preservation of signal details.
n 2
i=1 [(x(i ) − x(ˆi )]
RMSE = (10)
n
Fig. 8. An example of an experimental result waveform for eliminating noise from 4.2. Results and discussion
an EEG signal containing EMG noise of SNR = 0 dB. From top to bottom, the re-
sults are denoised by ICA, FICA, RLS filter, WT, DNN and the method proposed in
this paper. The dotted line represents the noisy EEG waveform, and the solid line To further verify the denoising performance of the proposed de-
represents the EEG waveform after using different denoising methods. noising method, it is compared with ICA [28], FICA [50], RLS filter
[51], WT [52] and DNN.
Fig. 11. Comparison of spectrograms: (a) the noisy EEG signal of the EOG noise at
0 dB, and the speech enhanced by (b) 1D-ResCNN, (c) DNN, (d) WT, (e) FICA, (f)
ICA, (g) RLS filter and (h) the clean EEG signal.
tion, the corrected EEG details are not preserved by these methods
compared to the original EEG.
Fig. 10 is the spectrum analysis of EOG noise removal for SNR =
0dB using various denoising methods. As the results show, the pro-
posed method and the noise-reduced DNN method are closer to
the clean EEG signal spectrum.
By removing ECG, EMG and EOG different noise waveform com-
parison graphs Figs. 6, 8 and 10 and spectrum analysis graphs
Figs. 7, 9, and 11, we can see that the proposed method performs
well for all three noise elimination, and DNN performs better for
the elimination of EOG noise.
Table 2
SNR experimental results of ECG noise with different SNR.
Evaluation index SNR −5dB −3dB −1dB 0dB 1dB 3dB 5dB
Table 3
SNR experimental results of EMG noise with different SNR.
Evaluation index SNR −5dB −3dB −1dB 0dB 1dB 3dB 5dB
Table 4
SNR experimental results of EOG noise with different SNR.
Evaluation index SNR −5dB −3dB −1dB 0dB 1dB 3dB 5dB
Table 5
Add RMSE experimental results of ECG noise.
Evaluation index RMSE −5dB −3dB −1dB 0dB 1dB 3dB 5dB
Table 6
Add RMSE experimental results of EMG noise.
Evaluation index RMSE −5dB −3dB −1dB 0dB 1dB 3dB 5dB
Table 7
Add RMSE experimental results of EOG noise.
Evaluation index RMSE −5dB −3dB −1dB 0dB 1dB 3dB 5dB
EEG information remains in the filtered EEG and these artifacts are U test [53] was used in this study. It assumes that the two
almost completely eliminated. samples are from the two identical populations except the
population mean, if the significance level is less than 0.05,
4.3.4. Hypothesis test then assume (two Overall equal) will be rejected (p value
In order to determine whether there is a significant differ- < 0.05). Tables 10–12 respectively show RMSE, SNR and ApEn
ence between the results of noise reduction by these meth- of Mann–Whitney U test result with different noise denoised
ods, the non-parametric statistical hypothesis test Mann–Whitney signals.
118 W. Sun, Y. Su and X. Wu et al. / Neurocomputing 404 (2020) 108–121
Table 8
Correlation between EEG signal and true EEG signal.
Evaluation index correlation ICA FICA RLS filter WT DNN Proposed method
Table 9
The ApEn comparison between Clean EEG and removed artifact from EEG signal.
Evaluation index ApEn ICA FICA RLS filter WT DNN Proposed method
Table 10
Hypothesis test result of ECG noise removel.
Table 11
Hypothesis test result of EMG noise removel.
Table 12
Hypothesis test result of EOG noise removel.
Fig. 14. The PSD result of eliminating ECG noise from the EEG signal, from the top
and bottom, by the ICA, FICA, RLS filter, WT, DNN and the algorithm proposed in
this paper, the EEG contaminated by the ECG, the EEG contaminated by the EMG,
by the EOG Contaminated EEG and corrected ESG PSD.
Fig. 15. Run-time bar graphs for different denoising methods ICA, FICA, RLS filter,
Fig. 13. RMSE results of remove noise from EEG signal. WT, DNN and the methods proposed in this paper.
Table 13 5. Conclusion
The average SNR for DNN and 1D-ResCNN methods.
Noise type SNR(dB) Noisy EEG DNN 1D-ResCNN In order to further improve the EEG denoising performance un-
GWN −10 −11.90 −0.69 −0.05 der unknown noise, CNN has better local feature expression abil-
0 −4.50 0.34 0.95 ity and can better utilize the correlation between EEG signal and
10 4.43 6.24 6.80 noise signal. In this paper, a reasonable network structure suiTable
EM −10 −12.21 −1.40 −0.96 for EEG denoising is proposed. An EEG denoising method based
0 −4.90 −0.39 0.11
on deep one-dimensional residual convolutional neural network is
10 3.97 5.53 7.03
proposed. The deep convolutional neural network is used to estab-
lish a regression model to express the complex nonlinear relation-
Table 14
ship between noisy EEG signals and pure EEG signals.Convolutional
The average RMSE for DNN and 1D-ResCNN methods.
networks filter residual data, use large data sets to improve sig-
Noise type SNR(dB) DNN 1D-ResCNN nal denoising, and help capture the main features of EEG sig-
GWN −10 0.4992 0.3847 nals. A large number of experimental results show that the one-
0 0.4820 0.3824 dimensional convolutional neural network proposed in this paper
10 0.4624 0.3718 achieves a smaller RMSE and better signal-to-noise ratio on the
EM −10 0.5503 0.5116
test set, and has better noise suppression ability than other de-
0 0.5416 0.5020
10 0.5371 0.4814 noising methods. Under various noise and various signal-to-noise
ratios, the nonlinear characteristics of EEG after denoising are sig-
nificantly maintained, and the EEG denoising performance under
unknown noise is further improved.
Acknowledgments
[13] A. Jafarifarmand, M.A. Badamchizadeh, Artifacts removal in eeg signal using a [41] Y.R. Tabar, U. Halici, A novel deep learning approach for classification of EEG
new neural network enhanced adaptive filter, Neurocomputing 103 (2) (2013) motor imagery signals, J. Neural Eng. 14 (1) (2016) 16003.
222–231. [42] T. Wen, Z. Zhang, Deep convolution neural network and autoencoder-
[14] J.E. Meng, Y. Gao, An adaptive RBFN-based filter for adaptive noise cancellation, s-based unsupervised feature learning of eeg signals, IEEE Access 6 (2018)
in: IEEE Conference on Decision and Control, 2003. Proceedings, 6, IEEE, 2003, 25399–25410.
pp. 6175–6180. [43] I. Ullah, M. Hussain, E.U.H. Qazi, H. Aboalsamh, An automated system for
[15] L. Fan, E.M. Joo, Learning algorithm for constructing fuzzy neural networks epilepsy detection using eeg brain signals based on deep learning approach,
with application to regression problems, in: International Conference on In- Expert Syst. Appl. 107 (2018) 61–71.
formation Science and Technology, 2011, pp. 318–322. [44] S. Ioffe, C. Szegedy, in: Batch Normalization: Accelerating Deep Network Train-
[16] C.K.S. Vijilal, P. Kanagasabapathy, S. Johnson, V. Ewards, Artifacts removal in ing by Reducing Internal Covariate Shift, 2015, pp. 448–456.
EEG signal using adaptive neuro fuzzy inference system, in: International [45] K. He, S. Jian, Convolutional neural networks at constrained time cost, in: Com-
Conference on Signal Processing, Communications and NETWORKING, 2007, puter Vision & Pattern Recognition, 2015.
pp. 589–591. [46] K. He, X. Zhang, S. Ren, J. Sun, in: Deep Residual Learning for Image Recog-
[17] J. Hu, C.S. Wang, M. Wu, Y.X. Du, Y. He, J. She, Removal of EOG and EMG arti- nition, In Proceedings of the IEEE conference on computer vision and pattern
facts from eeg using combination of functional link neural network and adap- recognition, 2016, pp. 770–778.
tive neural fuzzy inference system? Neurocomputing 151 (2015) 278–287. [47] S. Zagoruyko, N. Komodakis, in: Wide Residual Networks, 2016.
[18] M.K. Ahirwal, A. Kumar, G.K. Singh, Adaptive filtering of EEG/ERP through [48] Y. Pang, M. Sun, X. Jiang, X. Li, Convolution in convolution for network in net-
noise cancellers using an improved PSO algorithm, Swarm Evolut. Comput. 14 work, IEEE Trans. Neural Netw. Learn.Syst. 29 (5) (2017) 1587–1597.
(2014) 76–91. [49] C. Szegedy, V. Vanhoucke, S. Ioffe, J. Shlens, Z. Wojna, Rethinking the inception
[19] H.A.T. Nguyen, J. Musson, F. Li, W. Wang, G. Zhang, R. Xu, C. Richey, T. Schnell, architecture for computer vision, in: Computer Vision & Pattern Recognition,
F.D. Mckenzie, J. Li, EOG artifact removal using a wavelet neural network, Neu- 2016, pp. 2818–2826.
rocomputing 97 (1) (2012) 374–389. [50] A. Hyvarinen, Fast and robust fixed-point algorithms for independent compo-
[20] B. Noureddin, P.D. Lawrence, G.E. Birch, Time-frequency analysis of eye blinks nent analysis, IEEE Trans. Neural Netw. 10 (3) (1999) 626–634.
and saccades in EOG for EEG artifact removal, in: International IEEE/EMBS Con- [51] D.P. He, G. Wilson, C. Russell, Removal of ocular artifacts from electro-en-
ference on Neural Engineering, 2007, pp. 564–567. cephalogram by adaptive filtering, Med. Biol. Eng. Comput. 42 (3) (2004)
[21] N.E. Huang, Z. Shen, S.R. Long, M.C. Wu, H.H. Shih, Q. Zheng, N.-C. Yen, 407–412.
C.C. Tung, H.H. Liu, The empirical mode decomposition and the hilbert spec- [52] V. Krishnaveni, S. Jayaraman, L. Anitha, K. Ramadoss, Removal of ocular arti-
trum for nonlinear and non-stationary time series analysis, in: Proceedings of facts from eeg using adaptive thresholding of wavelet coefficients, J. Neural
the Royal Society of London A: Mathematical, Physical and Engineering Sci- Eng. 3 (4) (2006) 338.
ences, 454, The Royal Society, 1998, pp. 903–995. [53] M.P. Fay, M.A. Proschan, Wilcoxon-Mann–Whitney or t-test? On assumptions
[22] K.T. Sweeney, S.F. Mcloone, T.E. Ward, The use of ensemble empirical mode for hypothesis tests and multiple interpretations of decision rules, Stat. Surv. 4
decomposition with canonical correlation analysis as a novel artifact removal (2010) 1–39.
technique, IEEE Trans. Biomed. Eng. 60 (1) (2012) 97–105.
[23] A. Egambaram, N. Badruddin, V.S. Asirvadam, T. Begum, Comparison of enve- Weitong Sun received B.S. degree in 2014. Now, she is
lope interpolation techniques in empirical mode decomposition (EMD) for eye- pursuing Ph.D. Degree in Computer Software and Theory
blink artifact removal from EEG, in: Biomedical Engineering and Sciences, IEEE, from School of Computer Science, Shaanxi Normal Univer-
2016, pp. 590–595. sity. Her research interests include nonlinear dynamics,
[24] J. Li, C. Li, N. Thakor, A. Cichocki, A. Bezerianos, J. Li, C. Li, N. Thakor, chaotic systems, signal processing and machine learning.
A. Cichocki, A. Bezerianos, A new perspective of noise removal from EEG,
in: International IEEE/EMBS Conference on Neural Engineering, IEEE, 2017,
pp. 501–504.
[25] N. Siew-Cheok, R. Paramesran, Enhanced ${\mu} $ rhythm extraction using
blind source separation and wavelet transform, IEEE Trans. Biomed. Eng. 56
(8) (2009) 2024–2034.
[26] L. Shoker, S. Sanei, M.A. Latif, Removal of eye blinking artifacts from eeg in-
corporating a new constrained BSS algorithm, in: Sensor Array & Multichannel
Signal Processing Workshop, IEEE, 2004, pp. 177–181. Yuping Su, received the Ph.D. degree from the school
[27] K. Ting, P. Fung, C. Chang, F. Chan, Automatic correction of artifact from sin- of Communication Engeering, Xi’dian University, Xi’an,
gle-trial event-related potentials by blind source separation using second order China, in 2015.She is currently a lecturer of the School of
statistics only, Med. Eng. Phys. 28 (8) (2006) 780–794. Computer and Science, Shaanxi Normal University, Xi’an,
[28] Y. Li, Z. Ma, W. Lu, Y. Li, Automatic removal of the eye blink artifact from eeg China.Her research interests include Data mining and Ma-
using an ICA-based template matching approach, Physiol Meas 27 (4) (2006) chine Leaning.
425–436.
[29] T.-P. Jung, S. Makeig, A.J. Bell, T.J. Sejnowski, Independent component analysis
of electroencephalographic and event-related potential data, Adv. Neural Inf.
Process. Syst. 8 (8) (1996) 145–151.
[30] Y. Chen, Q. Zhao, B. Hu, J. Li, H. Jiang, W. Lin, Y. Li, S. Zhou, H. Peng, A
method of removing ocular artifacts from eeg using discrete wavelet transform
and Kalman filtering, in: IEEE International Conference on Bioinformatics and
Biomedicine, IEEE, 2016, pp. 1485–1492. Xia Wu received B.S. degree in Mathematics and Applied
[31] C. Burger, D.J.V.D. Heever, Removal of eog artefacts by combining wavelet Mathematics from college of Science, Xidian University
neural network and independent component analysis, Biomed. Signal Process. and M.S. degree from Key Laboratory of Modern Teach-
Control 15 (15) (2015) 67–79. ing technology, Ministry of Education, Shaanxi Normal
[32] P.S. Kumar, R. Arumuganathan, K. Sivakumar, C. Vimal, Removal of artifacts University. Now, she is pursuing Ph.D. Degree in Com-
from EEG signals using adaptive filter through wavelet transform, in: Inter- puter Software and Theory from School of Computer Sci-
national Conference on Signal Processing, IEEE, 2008, pp. 2138–2141. ence, Shaanxi Normal University. Her research interests
[33] Y. Xu, J. Du, L.R. Dai, C.H. Lee, An experimental study on speech enhancement include nonlinear dynamics, chaotic systems, signal pro-
based on deep neural networks, IEEE Signal Process. Lett. 21 (1) (2013) 65–68. cessing and machine learning.
[34] R. Rodrigues, P. Couto, A neural network approach to ECG denoising, Comput.
Sci. (2012).
[35] P. Xiong, H. Wang, M. Liu, S. Zhou, Z. Hou, X. Liu, ECG signal enhancement
based on improved denoising auto-encoder, Eng. Appl. Artif. Intell. 52 (C)
(2016) 194–202.
[36] B. Yang, K. Duan, C. Fan, C. Hu, J. Wang, Automatic ocular artifacts removal in WU Xiaojun (corresponding author)is a professor in
eeg using deep learning, Biomed. Signal Process. Control 43 (2018) 148–158. Shaanxi Normal University, Xi’an, China. His research in-
[37] Z. Zhang, F. Xing, H. Su, X. Shi, L. Yang, in: Recent Advances in the Applications terests include pattern recognition, intelligent system and
of Convolutional Neural Networks to Medical Image Contour Detection, 2017. system complexity.
[38] S. Bartunov, A. Santoro, B.A. Richards, G.E. Hinton, T. Lillicrap, in: Assess-
ing the Scalability of Biologically-motivated Deep Learning Algorithms and
Architectures, In Advances in Neural Information Processing Systems, 2018,
pp. 9368–9378.
[39] T. Jiang, Y. Zhang, X. Wu, G. Lu, F. Hao, Y. Zhang, Single image super resolution
via a refined densely connected inception network, in: 2018 25th IEEE Inter-
national Conference on Image Processing (ICIP), IEEE, 2018, pp. 3588–3592.
[40] P. Bashivan, I. Rish, M. Yeasin, N. Codella, Learning representations from EEG
with deep recurrent-convolutional neural networks, Comput. Sci. (2015).