Browsing by Author "Sen, Sena Yagmur"

Now showing 1 - 3 of 3

A Novel Face Identification Implementation for Class Attendance Monitoring
(Institute of Electrical and Electronics Engineers Inc., 2019) Hayriye Donmez; Sena Yagmur Sen; Nedim Orta; Atakan Aylanc; Ibrahim Zincir; Donmez, Hayriye; Sen, Sena Yagmur; Zincir, Ibrahim; Orta, Nedim; Aylanc, Atakan
Face identification has become more significant and relevant in the recent years. It is widely used for security purposes in enterprises and state-owned business since it has many advantages and benefits compared to other state of the art security applications. Previous face identification implementations inherited many different approaches and algorithms in order to overcome the challenges of recognizing an individual from a variety of angles and heights but none of them were completely successful. The main goal of this research is to demonstrate a novel face identification framework for an autonomous class attendance monitoring system implementing SIFT (Scale Invariant Feature Transform) algorithm. An image dataset generated with the participation of 20 volunteers that were photographed from a variety of different angles and heights was tested with the proposed system and achieved successful results in general with reasonable accuracy rates. © 2020 Elsevier B.V. All rights reserved.
Citation - Scopus: 78
Convolutional Neural Network Hyperparameter Tuning with Adam Optimizer for ECG Classification
(Institute of Electrical and Electronics Engineers Inc., 2020) Sena Yagmur Sen; Nalan Ǒzkurt; Sen, Sena Yagmur; Ozkurt, Nalan
In this research Adaptive Moment Estimation (Adam) optimization technique has been examined on ECG arrhythmia data that rely on deep neural networks. The proposed method indicates that Adam has great importance to solve deep learning problems. According to the proposed method the heartbeats are classified as normal (N) left bundle branch block (LBBB) and right bundle branch block (RBBB) considering the hyper-parameter tuning of the convolutional neural network (CNN). The heartbeats are transformed into spectrogram images and directly given into CNN without any feature extraction method but bounded with a specific frequency/time-resolution rate. The most important point of the study is the examination of the moment estimation coefficients of Adam optimizer such as first moment and second moments. Other tuned parameters are adaptive learning rate and epsilon value. The hyperparameters such as the learning rate and the moment estimation are investigated by grid search method. The effect of the parameters to validation loss were presented and analyzed as a result of this study. © 2020 Elsevier B.V. All rights reserved.
Citation - WoS: 11
Citation - Scopus: 27
ECG Arrhythmia Classification By Using Convolutional Neural Network And Spectrogram
(IEEE, 2019) Sena Yagmur Sen; Nalan Ozkurt; Sen, Sena Yagmur; Ozkurt, Nalan
In this study the electrocardiography (ECG) arrhythmias have been classified by the proposed framework depend on deep neural networks in order to features information. The proposed approaches operates with a large volume of raw ECG time-series data and ECG signal spectrograms as inputs to a deep convolutional neural networks (CNN). Heartbeats are classified as normal ( N) premature ventricular contractions (PVC) right bundle branch block (RBBB) rhythm by using ECG signals obtained from MIT-BIH arrhythmia database. The first approach is to directly use ECG time-series signals as input to CNN and in the second approach ECG signals are converted into time-frequency domain matrices and sent to CNN. The most appropriate parameters such as number of the layers size and number of the filters are optimized heuristically for fast and efficient operation of the CNN algorithm. The proposed system demonstrated high classification rate for the time-series data and spectrograms by using deep learning algorithms without standard feature extraction methods. Performance evaluation is based on the average sensitivity specificity and accuracy values. It is also worth to note that spectrogram increases the performance of classification since it extracts the useful time-frequency information of the signal.