Zümray Dokur
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Zümray Dokur has a total of 12 co-authors. Here are the top co-authors by number of shared publications:
- Tamer Ölmez
Istanbul Technical University
13 shared publications - Zafer Iscan
Centre for Cognition and Decision Making, National Research University Higher School of Economics, Russian Federation
4 shared publications - Ali Özkahraman
Department of Electronics and Communication Engineering, Istanbul Technical University, Maslak-Istanbul, Turkey
3 shared publications - Ayhan Yüksel
Department of Electronics and Communication Engineering, Istanbul Technical University, 34469 Maslak, Istanbul, Turkey
2 shared publications - Mehmet Korürek
Department of Electronics and Communication Engineering, Istanbul Technical University, 34469 Maslak, Istanbul, Turkey
2 shared publications - Nuri Korhan
İstanbul Technical University Maslak İstanbul Turkey
2 shared publications
Latest Publications
- 29 December 2024
- Published by Springer Nature in Neural Computing and Applications
- Vol. 37 (6) , 5057-5076
Determination of the common electrodes for users and increasing the classification accuracy of motor imagery EEG
- 29 December 2024
- Published by Springer Nature in Neural Computing and Applications
- Vol. 37 (6) , 5057-5076
Determination of the common electrodes for users and increasing the classification accuracy of motor imagery EEG
- Open Access
Performance Improvement with Reduced Number of Channels in Motor Imagery BCI System
Classifying Motor Imaging (MI) Electroencephalogram (EEG) signals is of vital importance for Brain–Computer Interface (BCI) systems, but challenges remain. A key challenge is to reduce the number of channels to improve flexibility, portability, and computational efficiency, especially in multi-class scenarios where more channels are needed for accurate classification. This study demonstrates that combining Electrooculogram (EOG) channels with a reduced set of EEG channels is more effective than relying on a large number of EEG channels alone. EOG channels provide useful information for MI signal classification, countering the notion that they only introduce eye-related noise. The study uses advanced deep learning techniques, including multiple 1D convolution blocks and depthwise-separable convolutions, to optimize classification accuracy. The findings in this study are tested on two datasets: dataset 1, the BCI Competition IV Dataset IIa (4-class MI), and dataset 2, the Weibo dataset (7-class MI). The performance for dataset 1, utilizing 3 EEG and 3 EOG channels (6 channels total), is of 83% accuracy, while dataset 2, with 3 EEG and 2 EOG channels (5 channels total), achieves an accuracy of 61%, demonstrating the effectiveness of the proposed channel reduction method and deep learning model.
Open AccessPerformance Improvement with Reduced Number of Channels in Motor Imagery BCI System
Classifying Motor Imaging (MI) Electroencephalogram (EEG) signals is of vital importance for Brain–Computer Interface (BCI) systems, but challenges remain. A key challenge is to reduce the number of channels to improve flexibility, portability, and computational efficiency, especially in multi-class scenarios where more channels are needed for accurate classification. This study demonstrates that combining Electrooculogram (EOG) channels with a reduced set of EEG channels is more effective than relying on a large number of EEG channels alone. EOG channels provide useful information for MI signal classification, countering the notion that they only introduce eye-related noise. The study uses advanced deep learning techniques, including multiple 1D convolution blocks and depthwise-separable convolutions, to optimize classification accuracy. The findings in this study are tested on two datasets: dataset 1, the BCI Competition IV Dataset IIa (4-class MI), and dataset 2, the Weibo dataset (7-class MI). The performance for dataset 1, utilizing 3 EEG and 3 EOG channels (6 channels total), is of 83% accuracy, while dataset 2, with 3 EEG and 2 EOG channels (5 channels total), achieves an accuracy of 61%, demonstrating the effectiveness of the proposed channel reduction method and deep learning model.
- 16 October 2024
- Published by Springer Nature
- p. 78-89
Impact of Noise Elimination Methods on Classification Performance in Motor Imagery EEG
- 16 October 2024
- Published by Springer Nature
- p. 78-89
Impact of Noise Elimination Methods on Classification Performance in Motor Imagery EEG
- 15 March 2023
- Published by Springer Nature
- p. 244-256
Comparison of Small-Sized Deep Neural Network Models for Hyperspectral Image Classification
- 15 March 2023
- Published by Springer Nature
- p. 244-256
Comparison of Small-Sized Deep Neural Network Models for Hyperspectral Image Classification
- Open Access
- 13 October 2022
- Published by Institute of Electrical and Electronics Engineers (IEEE) in TechRxiv
Hyperspectral Image Classification by A Novel 1D-DNN Trained with Walsh Functions
It is observed that high classification performances are achieved by using deep neural networks (DNNs). In this context, most researchers have employed these networks for the classification of hyperspectral images (HSIs). In HSI classification by the 2D/3D CNN, sub-windows with sizes larger than 11´11 are generally formed to move the filters of the CNNs. Thus, researchers turned to use spatial and spectral features jointly. The combination of spectral and spatial features causes some problems. Because the spatial resolution is not good in HSIs, these images usually do not contain strong textural information. In fact, the features that do not contribute significantly cause scattering of feature vectors in the input space. Another problem is that the training set implicitly includes data of the test set because of using sub-windows with sizes larger than 11´11, and high sampling rates. So, instead of using 2D/3D CNNs, we propose to use a novel 1D-DNN and train it by the Walsh functions. An average accuracy of 97% is achieved by using only the spectral data with one dimension and the novel 1D-DNN structure for the Indian Pines, Salinas, Pavia Centre, Pavia University and Botswana datasets. We have also investigated whether adding spatial data to the spectral data causes scattering in the input space, and calculated the percentage of the overlaps of the training and test sets with 11´11 sub-windows for less than 5% sampling rates.
- 13 October 2022
- Published by Institute of Electrical and Electronics Engineers (IEEE) in TechRxiv
Open AccessHyperspectral Image Classification by A Novel 1D-DNN Trained with Walsh Functions
It is observed that high classification performances are achieved by using deep neural networks (DNNs). In this context, most researchers have employed these networks for the classification of hyperspectral images (HSIs). In HSI classification by the 2D/3D CNN, sub-windows with sizes larger than 11´11 are generally formed to move the filters of the CNNs. Thus, researchers turned to use spatial and spectral features jointly. The combination of spectral and spatial features causes some problems. Because the spatial resolution is not good in HSIs, these images usually do not contain strong textural information. In fact, the features that do not contribute significantly cause scattering of feature vectors in the input space. Another problem is that the training set implicitly includes data of the test set because of using sub-windows with sizes larger than 11´11, and high sampling rates. So, instead of using 2D/3D CNNs, we propose to use a novel 1D-DNN and train it by the Walsh functions. An average accuracy of 97% is achieved by using only the spectral data with one dimension and the novel 1D-DNN structure for the Indian Pines, Salinas, Pavia Centre, Pavia University and Botswana datasets. We have also investigated whether adding spatial data to the spectral data causes scattering in the input space, and calculated the percentage of the overlaps of the training and test sets with 11´11 sub-windows for less than 5% sampling rates.
- 26 September 2022
- Published by Springer Nature in Neural Computing and Applications
- Vol. 35 (2) , 1303-1322
Generating ten BCI commands using four simple motor imageries and classification by divergence-based DNN
- 26 September 2022
- Published by Springer Nature in Neural Computing and Applications
- Vol. 35 (2) , 1303-1322
Generating ten BCI commands using four simple motor imageries and classification by divergence-based DNN
- Open Access
- 31 December 2021
- Published by International Journal of Applied Mathematics, Electronics and Computers in International Journal of Applied Mathematics Electronics and Computers
- Vol. 9 (4) , 85-90
Classification of left and right hand motor imagery EEG signals by using deep neural networks
The brain-computer interface (BCI) is one of the most promising technologies that allows us to establish a relationship between brain and devices. In this study, three-channel EEG signals collected from nine subjects performing two motor imagery tasks are classified using two different deep neural network (DNN) based approaches called framework 1 (FW1) and framework 2 (FW2). The proposed frameworks were evaluated using BCI Competition IV-IIb dataset. In FW1, the raw EEG data is directly presented to the deep neural network without performing any pre-processing. In FW2, the EEG data is first filtered with five band pass filters with fifth order (Butterworth), then the common spatial patterns (CSP) method, which introduces additional pseudo channels, is applied to the filtered signals. Two experiments were conducted for each framework. In the first experiment, a unique DNN is trained for each subject, and in the second experiment only one DNN is trained with the combination of training sets of all subjects. The performance of the two experiments are then compared in terms of average accuracy. According to the simulation results, we did not observe a significant difference between the average classification accuracies obtained with the first and the second experiments. Therefore, we concluded that, by the use of DNNs we do not need to train several subject-specific networks which requires high computational loads. On the other hand, we observed that the average classification performance significantly improves by the filtering and extracting features with CSP pre-processes.
- 31 December 2021
- Published by International Journal of Applied Mathematics, Electronics and Computers in International Journal of Applied Mathematics Electronics and Computers
- Vol. 9 (4) , 85-90
Open AccessClassification of left and right hand motor imagery EEG signals by using deep neural networks
The brain-computer interface (BCI) is one of the most promising technologies that allows us to establish a relationship between brain and devices. In this study, three-channel EEG signals collected from nine subjects performing two motor imagery tasks are classified using two different deep neural network (DNN) based approaches called framework 1 (FW1) and framework 2 (FW2). The proposed frameworks were evaluated using BCI Competition IV-IIb dataset. In FW1, the raw EEG data is directly presented to the deep neural network without performing any pre-processing. In FW2, the EEG data is first filtered with five band pass filters with fifth order (Butterworth), then the common spatial patterns (CSP) method, which introduces additional pseudo channels, is applied to the filtered signals. Two experiments were conducted for each framework. In the first experiment, a unique DNN is trained for each subject, and in the second experiment only one DNN is trained with the combination of training sets of all subjects. The performance of the two experiments are then compared in terms of average accuracy. According to the simulation results, we did not observe a significant difference between the average classification accuracies obtained with the first and the second experiments. Therefore, we concluded that, by the use of DNNs we do not need to train several subject-specific networks which requires high computational loads. On the other hand, we observed that the average classification performance significantly improves by the filtering and extracting features with CSP pre-processes.
- 16 January 2020
- Published by Springer Nature in Neural Computing and Applications
- Vol. 32 (16) , 12515-12534
Heartbeat classification by using a convolutional neural network trained with Walsh functions
- 16 January 2020
- Published by Springer Nature in Neural Computing and Applications
- Vol. 32 (16) , 12515-12534
Heartbeat classification by using a convolutional neural network trained with Walsh functions