Browsing by Author "Toprak, Tugce"

Now showing 1 - 5 of 5

Citation - WoS: 27
Citation - Scopus: 34
Conditional Weighted Ensemble of Transferred Models for Camera Based Onboard Pedestrian Detection in Railway Driver Support Systems
(IEEE-INST ELECTRICAL ELECTRONICS ENGINEERS INC, 2020) Tugce Toprak; Burak Belenlioglu; Burak Aydin; Cuneyt Guzelis; M. Alper Selver; Toprak, Tugce; Guzelis, Cuneyt; Selver, M. Alper; Belenlioglu, Burak; Aydin, Burak
Pedestrian Detection (PD) is one of the most studied issues of driver assistance systems. Although a tremendous effort is already given to create datasets and to develop classifiers for cars studies about railway systems remain very limited. This article shows that direct application of neither existing advanced object detectors (such as AlexNet VGG YOLOetc.) nor specifically created systems for PD(such as Caltech/INRIA trained classifiers) can provide enough performance to overcome railway specific challenges. Fortunately it is also shown that without waiting the collection of a mature dataset for railways as comprehensively diverse and annotated as the existing ones for cars a Transfer Learning (TL) approach to fine-tune various successful deep models (pre-trained using both extensive image and pedestrian datasets) to railway PD tasks provides an effective solution. To achieve TL a new RAil-Way PEdestrian Dataset (RAWPED) is collected and annotated. Then a novel three-stage system is designed. At its first stage a feature-classifier fusion is created to overcome the localization and adaptation limitations of deep models. At the second stage the complementarity of the transferred models and diversity of their results are exploited by conducted measurements and analyses. Based on the findings at the third stage a novel learning strategy is developed to create an ensemble which conditionally weights the outputs of individual models and performs consistently better than its components. The proposed system is shown to achieve a log average miss rate of 0.34 and average precision of 0.93 which are significantly better than the performance of compared well-established models.
Searching Optimal Values of Identification and Controller Design Horizon Lengths and Regularization Parameters in NARMA Based Online Learning Controller Design
(Institute of Electrical and Electronics Engineers Inc., 2019) Tugce Toprak; Savaş Şahin; Mehmet Uğur Soydemir; Parvin Bulucu; Aykut Kocaoǧlu; Cüneyt Güzeliş; Toprak, Tugce; Bulucu, Parvin; Kocaoglu, Aykut; Soydemir, M. Ugur; Guzelis, Cuneyt; Sahin, Savas
This paper presents an analysis on searching the optimal values of the system identification and tracking window lengths and regularization parameter for the online learning NARMA controller algorithm. Both window lengths and regularization parameter are generally determined with exhaustive searches by researchers. Although the estimation of plant and controller parameters plays the essential role in online learning control algorithms using non-optimal values of the window lengths and regularization parameter may deteriorate badly the estimation and so the performance of the controller. In the paper the effects of the window lengths and the regularization parameter on the tracking performance of the NARMA based online learning controller are analyzed with a search method. The considered NARMA based online learning control method is performed on a rotary inverted pendulum model. While the effect of the regularization parameter is examined in the batch mode the effects of identification and tracking error window lengths are studied for the online mode of the controller learning algorithm. The developed search method can provide the optimum values of the plant identification and tracking horizon lengths and regularization parameter when a sufficiently large class of possible input output and reference signals are taken into account in the search. The presented study may be extended as future research in the direction of developing intelligent control systems by determining the horizon window lengths and regularization parameter in an automatic way with efficient learning algorithms. © 2020 Elsevier B.V. All rights reserved.
Citation - WoS: 5
Citation - Scopus: 5
Transferring Synthetic Elementary Learning Tasks to Classification of Complex Targets
(Institute of Electrical and Electronics Engineers Inc., 2019) Alper Mustafa Selver; Tugce Toprak; Mustafa Seçmen; Emine Yeşim Zoral; Toprak, Tugce; Zoral, E. Yesim; Selver, M. Alper; Secmen, Mustafa
Deep learning has a promising impact on target classification performance at the expense of huge training data requirements. Therefore the use of simulated data is inevitable for convergence of deep models (DMs). However generating synthetic data for real-life complex targets can be quite tedious and is not always possible. In this study DMs trained with synthetic one-dimensional scattered data of elementary targets are transferred to classify complex targets from measured signals for the first time. For this purpose a novel system is proposed by combining three strategies: first initial training of DMs using analytical and simulated time domain scattered data obtained from the basic targets, second the last layers of initial DMs are fine-tuned by transfer learning using measured signals of the real targets, and third an ensemble model is developed to generate a model that can completely represent real target characteristics by combining diverse and complementary properties of the fine-tuned DMs. The proposed system provides higher accuracy sensitivity and specificity performances compared to the existing methods. © 2019 Elsevier B.V. All rights reserved.
Citation - WoS: 1
Citation - Scopus: 1
Utilizing Resonant Scattering Signal Characteristics of Magnetic Spheres via Deep Learning for Improved Target Classification
(IEEE, 2019) M. Alper Selver; Tugce Toprak; Mustafa Secmen; E. Yesim Zoral; Toprak, Tugce; Zoral, E. Yesim; Selver, M. Alper; Secmen, Mustafa
Object classification using LAte-time Resonant Scattering Electromagnetic Signals (LARSESs) is a significant problem found in different areas of application. Due to their special properties spherical objects play an important role in this field both as a challenging target and analytical LARSES source. Although many studies focus on their detailed analysis the challenges associated with target classification by resonant LARSESs from multi-layer spheres have not been investigated in detail. Moreover existing studies made the simplifying assumption that the objects having (one or more) layers constitute equal permeability values at the core and coatings. However especially for metamaterials magneto-dielectric inclusions require consideration of magnetic properties as well as dielectric ones. In this respect this study shows that the utilization LARSESs of magnetic spheres provides diverse information and features which result with superior object classification performance. For this purpose first time domain LARSESs are generated numerically for single and multi-layer radially symmetrical dielectric and magnetic spheres. Then by using emerging deep learning tools particularly Convolutional Neural Network (CNNs) which are trained with spheres having different material properties a high multi-layer object classification performance is achieved. Moreover by extending the proposed strategy to measured data via modern data augmentation and transfer learning techniques an improved classification performance is also obtained for more complex targets.
Citation - WoS: 2
Citation - Scopus: 2
Utilizing resonant scattering signal characteristics via deep learning for improved classification of complex targets
(Turkiye Klinikleri, 2021) Tugce Toprak; Alper Mustafa Selver; Mustafa Seçmen; E. Yesim Zoral; Toprak, Tugce; Zoral, E. Yesim; Yesim Zoral, E.; Alper Selver, M.; Selver, M. Alper; Secmen, Mustafa
Object classification using late-time resonant scattering electromagnetic signals is a significant problem found in different areas of application. Due to their unique properties spherical objects play an essential role in this field both as a challenging target and a resource of analytical late-time resonant scattering electromagnetic signals. Although many studies focus on their detailed analysis the challenges associated with target classification by resonant late-time resonant scattering electromagnetic signals from multilayer spheres have not been investigated in detail. Moreover existing studies made the simplifying assumption that the objects having (one or more) layers constitute equal permeability values at the core and coatings. However especially for metamaterials magneto-dielectric inclusions require consideration of magnetic properties as well as dielectric ones. In this respect this study shows that the utilization late-time resonant scattering electromagnetic signals of magnetic spheres provide diverse information and features which result in superior object classification performance. For this purpose first time-domain late-time resonant scattering electromagnetic signals are generated numerically for single and multilayer radially symmetrical dielectric and magnetic spheres. Then by using emerging deep learning tools particularly convolutional neural networks trained with spheres having different material properties a high multilayer object classification performance is achieved. Furthermore by incorporating the frequency characteristics of the late-time resonant scattering electromagnetic signals to the classification process through Fourier transform and convolutional neural network layers for feature extraction a convolutional neural network with long short term memory algorithm is developed. The outcome of the proposed algorithm design is shown to be particularly successful even in the case of limited available data on challenging targets. This extended strategy is also shown to outperform modern data augmentation and transfer learning techniques in terms of accuracy as well as the computational cost. © 2021 Elsevier B.V. All rights reserved.