Browsing by Author "Özmen, Atilla"

Now showing 1 - 20 of 36

2.5d Vit: 3 Boyutlu Beyin Mr Görüntülerinin Ön İşlenmesiyle Görüntü Dönüştürücü Tabanlı Beyin Yaşı Tahmini
(2024) Darıcı, Muazzez Buket; Özmen, Atilla
Son zamanlarda doğal görüntü işleme görevinde kullanılan transformörler, görme görevlerine alternatif bir çözüm sunmaktadır. Görüntülerin işlenmesine olanak sağlayan görüntü transformör mimarisinin, güçlü dikkat mekanizması ve konumsal bilgiyi tutma yeteneği ile görüntü sınıflandırma görevinde etkili olduğu kanıtlanmıştır. Görüntü sınıflandırmaya yenilikçi bir yaklaşım olan ViT, popüler veri setlerinde güncel CNN'lerden daha iyi performans göstermektedir. Ne yazık ki ViT yapısı 2D ile uyumlu olduğundan, saf haliyle 2 boyuttan fazla olan görüntüleri işleyemez. Bu çalışma, 3 boyutlu beyin MR görüntülerini işleyebilen 2.5D ViT adlı yeni bir ViT önermektedir. Model mimarisinde yapılan değişiklikler ve önerilen yöntemler sayesinde 2.5D ViT, 3D görüntülerden yaş tahminini güncel modellere göre daha iyi yapabilmektedir. Ayrıca bu çalışma, beyin MR görüntülerinin hem model mimarisi hem de ön işleme aşamaları üzerine geniş çaplı deneyler içermektedir. Üstün başarısıyla insanların hayatına etki eden Yapay Zeka tabanlı beyin analiz sistemleri, ideal 3 boyutlu beyin MR görüntülerine ihtiyaç duyar. Bu sistemler için ideal beyin MR görüntüleri elde etmek amacıyla en çok tercih edilen ön işleme teknikleri Yanlılık Alanı Düzeltme (Bias Field Correction), Kafatası Sıyırma (Skull Stripping) ve Çakıştırmadır (Registration). Ön işlemin görüntüleri standartlaştırdığı bilinse bile, ön işlemlerin son teknolojiye sahip ağlarda beyin yaşı tahmin sistemlerinin kalitesi üzerindeki etkisi titizlikle araştırılmamıştır. Bu çalışma, IXI veri setinden alınan 3 boyutlu beyin MR görüntüleri üzerindeki Yanlılık Alanı Düzeltme ve Kafatası Sıyırma etkilerinin yanı sıra Çakıştırma sırasında uygulanan ön işlemlerin etkilerinin ve bunların sırasının kapsamlı bir şekilde gözlemlenmesini içermektedir. Beyin yaşı tahmini alanında popüler olan 3 boyutlu Evrişimsel Sinir Ağları modeli, ön işlemlerin beyin yaşı tahmini üzerindeki başarısı hakkında bilgi vermesi için kullanılmıştır. Bu çalışmanın çıktıları, ön işleme yöntemleri olarak sırasıyla Kafatası Sıyırma, Yanlılık Alanı Düzeltme, Çakıştırma işlemleri Z-Score normalizasyonu ile kullanıldığında, 3 boyutlu Evrişimsel Sinir Ağının 6 yıllık ortalama mutlak hata ile farklı şekilde önceden işlenmiş görüntüler üzerinde eğitilen diğer modellerden daha iyi performans gösterdiğini ortaya koymaktadır. Bu çalışmayı önemli kılan bir diğer nokta ise beyin yaşı tahmini üzerinde kullanıma hazır SPM aracına benzer performans gösterebilecek ön işleme tekniklerini uygun sırayla önermesidir. Önerilen tekniklerle önceden işlenmiş 3 boyutlu beyin MR görüntüleri daha sonra yeni Görüntü Dönüştürücü (ViT) için girdi olarak kullanılmıştır. 2.5D ViT'in tasarımı, beyin yaşı tahmin performansını maksimuma çıkarırken bilgi kaybını en aza indirmeye odaklanır. 2.5D ViT tasarımı ViT'den farklı olarak SCA'dan RGB'ye dönüşüm mimarisi ve Ayrık Kosinüs Dönüşümü (AKD) içermektedir. SCA'dan RGB'ye dönüşüm, 3 boyutlu görüntülerin maksimum bilgiyle 2 boyutlu görüntülere dönüştürülmesini sağlarken, güçlü sıkıştırma kabiliyetine sahip AKD, ViT'deki Dönüştürücü kodlayıcıyı besleyen, yaşa bağlı özellikleri içeren, daha küçük boyutta özellik haritası elde etmek için kullanılır. Çeşitli deneylerden sonra 2.5D ViT, yanlılık düzeltmesinden sonra 5 yıl mutlak hata oranı ile en iyi performansı elde etmektedir. Sonuçlar, önerilen 2.5D ViT'nin beyin yaşı tahmini alanında 3 boyutlu Evrişimsel Sinir Ağları ile karşılaştırmalı sonuçlara sahip olduğunu göstermektedir. Mutlak ortalama hataya ek olarak araştırılan istatistiksel değerler ise sırasıyla r değeri için 0.9, Spearman Korelasyon Katsayısı için 0.87 ve R Kare değeri ise ortalamada 0.78 olarak bulunmuştur. Bu değerler, yanlılık düzeltme işleminden sonraki değerlerdir.
Air Quality Prediction Using a Hybrid Deep Learning Architecture
(Kadir Has Üniversitesi, 2020) Gilik, Ayşenur; Özmen, Atilla
Air pollution prediction is related to the variables in environmental monitoring data and modeling of the complex relationship between these variables. The objectives of the thesis are to develop a supervised model for the prediction of air pollution by using real sensor data and to transfer the model between cities. A CNN+LSTM deep neural network model was developed to predict the concentration of air pollutants in multiple locations by using a spatial-temporal relationship. The 2D input (univariate) contains the information of one pollutant; the 3D input (multivariate) contains the information of all pollutants and meteorology. There are three methods employed according to the input-output type: Method-1 is based on univariate-input and univariate-output; Method-2 is based on multivariate input and univariate-output; Method-3 is based on multivariate input and multivariate output. The study was carried out for different pollutants which are in publicly available data of the cities of Barcelona, Kocaeli, and İstanbul. The hyperparameters were tuned to determine the architecture that achieved the lowest test RMSE. Comparing the performance of the CNN+LSTM network with a 1-hidden layer LSTM network, the proposed model improved the prediction performance by the rates between 11%-53% for PM10, 20%-31% for O3, 9%-47% for NOX and 18%-46% for SO2. After, the network weights were transferred from the source domains to the target domain. The model has a more reliable prediction performance with the transfer of the network from Kocaeli to İstanbul because of the similarities between those two cities.
Amplitude and Frequency Modulations With Cellular Neural Networks
(Springer, 2015) Tander, Baran; Özmen, Atilla
Amplitude and frequency modulations are still the most popular modulation techniques in data transmission at telecommunication systems such as radio and television broadcasting gsm etc. However the architectures of these individual systems are totally different. In this paper it is shown that a cellular neural network with an opposite-sign template can behave either as an amplitude or a frequency modulator. Firstly a brief information about these networks is given and then the amplitude and frequency surfaces of the generated quasi-sine oscillations are sketched with respect to various values of their cloning templates. Secondly it is proved that any of these types of modulations can be performed by only varying the template components without ever changing their structure. Finally a circuit is designed simulations are presented and performance of the proposed system is evaluated. The main contribution of this work is to show that both amplitude and frequency modulations can be realized under the same architecture with a simple technique specifically by treating the input signals as template components.
Analytical Approaches for the Amplitude and Frequency Computations in the Astable Cellular Neural Networks With Opposite Sign Templates
(IEEE, 2007) Tander, Baran; Özmen, Atilla
In this paper, by using surface fitting methods, analytical approaches for amplitudes and frequencies of the x(1,2)(t) "States" in a simple dynamical neural network called "Cellular Neural Network with Opposite Sign Templates" which was proposed by Zou and Nossek [1], are obtained under oscillation conditions. The mentioned explicit expressions are employed in a cellular neural network based, amplitude and frequency tuneable oscillator design.
Citation - WoS: 7
Citation - Scopus: 10
Artificial Neural Network Based Estimation of Sparse Multipath Channels in Ofdm Systems
(SPRINGER, VAN GODEWIJCKSTRAAT 30, 3311 GZ DORDRECHT, NETHERLANDS, 2021) Şenol, Habib; Abdur Rehman Bin, Tahir; Özmen, Atilla
In order to increase the transceiver performance in frequency selective fading channel environment, orthogonal frequency division multiplexing (OFDM) system is used to combat inter-symbol-interference. In this work, a channel estimation scheme for an OFDM system in the presence of sparse multipath channel is studied using the artificial neural networks (ANN). By means of ANN's learning capability, it is shown that how to model and obtain a channel estimate and how it allows the proposed technique to give a better system throughput. The performance of proposed method is compared with the Matching Pursuit (MP) and Orthogonal MP (OMP) algorithms that are commonly used in compressed sensing literature in order to estimate delay locations and tap coefficients of a sparse multipath channel. In this work, we propose a performance- efficient ANN based sparse channel estimator with lower computational cost than that of MP and OMP based channel estimators. Even though there is a slight performance lost in a few simulation scenarios in which we have lower computational complexity advantage, in most scenarios, our computer simulations corroborate that our low complexity ANN based channel estimator has better mean squared error and the corresponding symbol error rate performances comparing with MP and OMP algorithms.
Artificial Neural Network Based Sparse Channel Estimation for Ofdm Systems
(Kadir Has Üniversitesi, 2017) Tahir, Abdur Rehman Bin; Şenol, Habib; Özmen, Atilla
In order to increase the communication quality in frequency selective fading channel environment, orthogonal frequency division multiplexing (OFDM) systems are used to combat inter-symbol-interference (ISI). In this thesis, a channel estimation scheme for the OFDM system in the presence of sparse multipath channel is studied. The channel estimation is done by using the artificial neural networks (ANNs) with Resilient Backpropagation training algorithm. This technique uses the learning capability of artificial neural networks. By means of this feature we show how to obtain a channel estimate and how it allows the proposed technique to be less computationally complex; as there is no need for any matrix inversions. This proposed method is compared with the Matching Pursuit (MP) algorithm that is well known estimation technique for sparse channels. The results show that the ANN based channel estimate is computationally simpler and a small number of pilots are required to get a better estimate of the channel especially in low SNR levels. With this setting, the proposed algorithm leads to a better system throughput.
Autonomous Vehicle Control Using Reinforcement Learning
(Kadir Has Üniversitesi, 2020) Bozkurt, Hüma; Özmen, Atilla
Autonomous vehicles have become an important research topic where artificial intelligence is applied. As the research increases, by means of the applications of artificial intelligence algorithms in different areas, enable the working mechanisms of the systems to become more optimal due to the change of factors such as human power, time, energy and control. It has been observed that deep learning and machine learning algorithms have advantages and disadvantages in different situations and conditions. Since deep learning algorithms require large amounts of data, studies on the reinforcement learning model based on the experience from the environment and based on the reward-punishment system have recently concentrated and some striking results have been obtained. Reinforcement learning is considered a powerful AI paradigm that can be used to teach machines through interaction with the environment and learning from their mistakes. In this thesis, an environment was created based on a two-dimensional vehicle scenario created using a pyglet simulation tool. A comparative simulation study of different reinforcement learning algorithms such as Q-Learning, SARSA and Deep Q-Network (DQN) is presented on this environment. While making this comparison, a certain learning criterion was added, and also, parameters such as epsilon value, step number were changed, and changes in training and test stages were analyzed. For this study, the actors (agent, sensor, obstacles etc.) provided by the simulator program were supported. Through the feedback provided by the sensors, the reinforcement learning agent trains himself on the basis of these algorithms and determines a movement strategy to explore the environment limited to a specific area.
Citation - WoS: 1
Citation - Scopus: 1
Bayesian Estimation of Discrete-Time Cellular Neural Network Coefficients
(TUBITAK Scientific & Technical Research Council Turkey, 2017) Özer, Hakan Metin; Özmen, Atilla; Şenol, Habib
A new method for finding the network coefficients of a discrete-time cellular neural network (DTCNN) is proposed. This new method uses a probabilistic approach that itself uses Bayesian learning to estimate the network coefficients. A posterior probability density function (PDF) is composed using the likelihood and prior PDFs derived from the system model and prior information respectively. This posterior PDF is used to draw samples with the help of the Metropolis algorithm a special case of the Metropolis--Hastings algorithm where the proposal distribution function is symmetric and resulting samples are then averaged to find the minimum mean square error (MMSE) estimate of the network coefficients. A couple of image processing applications are performed using these estimated parameters and the results are compared with those of some well-known methods.
Caching Algorithm Implementation for Edge Computing in Iot Network
(Kadir Has Üniversitesi, 2020) Abduljabbar, Mohammed; Özmen, Atilla; Öğrenci, Arif Selçuk
The developing IoT concept brings new challenges to the service providers. The architecture of the networks changes to satisfy the needs arising by the large number of connected devices. Edge computing is the new architectural solution that will be used in the IoT networks. This architecture is more dynamic than the cloud computing network where the data can be quickly processed in the different layers of the network without going to the cloud. This will remove the problems faced by cloud computing: increase in data traffic and increase in latency of provided services. Research on edge computing in IoT networks encompass information-centric networks, use of 5G, and improving the hardware devices however a suitable solution for all the IoT use cases is not available yet. In this thesis, use of caching among IoT nodes is proposed as a solution to increase the efficiency of edge computing. Caching is an old but effective solution for dealing with data because it improves the real-time response of the system and can be used in IoT use cases. It will also not cause an extra hardware cost. In this research, two commonly used caching algorithms, LRU (Least Recently Used) and FIFO (First in First Out), are investigated and compared for their performance in sample IoT scenarios. Reductions in data processing time are observed where CPU and RAM utilizations are enhanced.
Citation - Scopus: 1
Channel Equalization With Cellular Neural Networks
(IEEE, 2010) Özmen, Atilla; Tander, Baran
In this paper a dynamic neural network structure called Cellular Neural Network (CNN) is employed for the equalization in digital communication. It is shown that this nonlinear system is capable of suppressing the effect of intersymbol interference (ISI) and the noise at the channel. The architecture is a small-scaled simple CNN containing 9 neurons thus having only 19 weight coefficients. Proposed system is compared with linear transversal filters as well as with a Multilayer Perceptron (MLP) based equalizer.
Citation - WoS: 4
Citation - Scopus: 6
Channel Estimation for Realistic Indoor Optical Wireless Communication in Aco-Ofdm Systems
(Springer, 2018) Özmen, Atilla; Şenol, Habib
In this paper channel estimation problem in a visible light communication system is considered. The information data is transmitted using asymmetrical clipped optical orthogonal frequency division multiplexing. Channel estimation and symbol detection are performed by the Maximum Likelihood and the Linear Minimum Mean Square Error detection techniques respectively. The system performance is investigated in realistic environment that is simulated using an indoor channel model. Two different channels are produced using the indoor channel model. Symbol error rate (SER) performance of the system with estimated channels is presented for QPSK and 16-QAM digital modulation types and compared with the perfect channel state information. As a mean square error (MSE) performance benchmark for the channel estimator Cramer-Rao lower bound is also derived. MSE and SER performances of the simulation results are presented.
Computation of Two-Variable Mixed Element Network Functions
(Kadir Has Üniversitesi, 2017) Tabassum, Nauman; Özmen, Atilla
in this dissertation the algorithm known as “Standard Decomposition Technique (SDT)” is used together with Belevitch’s canonic representation of scattering polynomial for two-port networks operate on high frequency to find the analytical solutions for “Fundamental equation set (FES)”. This FES is extracted by using Belevitch canonic polynomials “ ??(?? ??) ?(?? ??) and ??(?? ??)” used for the description of mixed lumped and distributed lossless two-port cascaded networks in two variables of degree five and the obtained solutions are further used to synthesis the realizable networks. The solution to the problem is also classified into two cases first case is discussed for three lumped and two distributed (???? = 3 ???? = 2 ) and the second is for three distributed and two lumped important (???? = 2 ???? = 3 ) the solution for both these cases are expressed separately with conclusive examples
Citation - WoS: 16
Citation - Scopus: 17
Correlation of Experimental Liquid-Liquid Equilibrium Data for Ternary Systems Using Nrtl and Gmdh-Type Neural Network
(Amer Chemical Soc, 2017) Bekri, Sezin; Özmen, Dilek; Özmen, Atilla
In this work liquid liquid equilibrium (LLE) data for the ternary systems (water + propionic acid + solvent) were experimentally obtained at atmospheric pressure and 298.2 K. The ternary systems show type-1 behavior of LLE. Cyclopentane cyclopentanol 2-octanone and dibutyl maleate were chosen as solvent and it has been noted that there are no data in the literature on these ternary systems. The consistency of the experimental tie-line data was checked using the Hand and Othrner-Tobias correlation equations. A comparison of the extracting capabilities of the solvent was made with respect to the distribution coefficients and separation factors. The correlation of the experimental tie-line data was confirmed by the NRTL thermodynamic model. A Group Method of Data Handling (GMDH)-type neural network (NN) was also used to correlate the experimental tie-lines. It is shown that the results of the both models cohere with the experimental values.
Citation - WoS: 6
Citation - Scopus: 6
Correlation of Ternary Liquid- Equilibrium Data Using Neural Network-Based Activity Coefficient Model
(Springer, 2014) Özmen, Atilla
Liquid--liquid equilibrium (LLE) data are important in chemical industry for the design of separation equipments and it is troublesome to determine experimentally. In this paper a new method for correlation of ternary LLE data is presented. The method is implemented by using a combined structure that uses genetic algorithm (GA)--trained neural network (NN). NN coefficients that satisfy the criterion of equilibrium were obtained by using GA. At the training phase experimental concentration data and corresponding activity coefficients were used as input and output respectively. At the test phase trained NN was used to correlate the whole experimental data by giving only one initial value. Calculated results were compared with the experimental data and very low root-mean-square deviation error values are obtained between experimental and calculated data. By using this model tie-line and solubility curve data of LLE can be obtained with only a few experimental data.
Citation - WoS: 1
Design and Implementation of a Cellular Neural Network Based Oscillator Circuit
(World Scientific and Engineering Acad and Soc, 2009) Tander, Baran; Özmen, Atilla; Özçelep, Yasin
In this paper, a novel inductorless oscillator circuit with negative feedbacks, based on a simple version of a "Cellular Neural Network" (CNN) called "CNN with an Opposite Sign Template" (CNN-OST) is designed and implemented. The system is capable of generating quasi-sine oscillations with tuneable amplitude and frequency which can't be provided at the same time in the conventional oscillator circuits.
Citation - Scopus: 5
Design and Implementation of a Negative Feedback Oscillator Circuit Based on a Cellular Neural Network With an Opposite Sign Template
(2010) Tander, Baran; Özmen, Atilla; Özçelep, Yasin
In this paper explicit amplitude and frequency expressions for a Cellular Neural Network with an Opposite-Sign Template (CNN-OST) under oscillation condition are derived and a novel inductorless oscillator circuit with negative feedbacks based on this simple structure is designed and implemented. The system is capable of generating quasi-sine signals with tuneable amplitude and frequency which can't be provided at the same time in the classical oscillator circuits.
Design of Low-Pass Ladder Networks With Mixed Lumped and Distributed Elements by Means of Artificial Neural Networks
(AVES YAYINCILIK, 2003) Şengül, Metin Y.; Özmen, Atilla; Yılmaz, Melek
In this paper, calculation of parameters of low-pass ladder networks with mixed lumped and distributed elements by means of artificial neural networks is given. The results of ANN are compared with the values that are desired. It has been observed that the calculated and the desired values are extremely close to each other. So this algorith can be used to obtain the parameters that will be used to synthesize such circuits.
Citation - Scopus: 3
Detection of Trojans in Integrated Circuits
(IEEE, 2012) Baktır, Selçuk; Güçlüoğlu, Tansal; Özmen, Atilla; Alsan, Hüseyin Fuat; Macit, Mustafa Can
This paper presents several signal processing approaches in Trojan detection problem in very large scale integrated circuits. Specifically wavelet transforms spectrograms and neural networks are used to analyze power side-channel signals. Trojans in integrated circuits can try to hide themselves and become almost invisible due to process and measurement noises. We demonstrate that our initial results with these techniques are promising in successful detection. Discrete wavelet transforms and spectrograms can provide clear visual assistance in detecting Trojans by catching the time-scale differences and time-frequency activities introduced by the Trojans. Furthermore neural networks with sufficient training are also used and simulation results show that correct decisions are possible with a very high success rate. © 2012 IEEE.
Citation - Scopus: 1
Edge Detection Using Steerable Filters and Cnn
(European Signal Processing Conference EUSIPCO, 2002) Özmen, Atilla; Akman, Emir Tufan
This paper proposes a new approach for edge detection using steerable filters and cellular neural networks (CNNs) where the former yields the local direction of dominant orientation and the latter provides iterative filtering. For this purpose steerable filter coefficients are used in CNN as a B template. The results are compared to the results where only CNN or steerable filters are used. As a result of this study the performance of the system can be improved since iterative filtering property of CNN and the ability of steerable filters for edge detection are used. © 2002 EUSIPCO.
The Effect of Data Augmentation on Adhd Diagnostic Model Using Deep Learning
(IEEE, 2019) Çiçek, Gülay; Özmen, Atilla; Akan, Aydin
Attention Deficit Hyperactivity Disorder (ADHD) is a neuro-behavioral hyperactivity disorder. It is frequently seen in childhood and youth, and lasts a lifetime unless treated.The ADHD classification model should be objective and robust. Correct diagnosis usually depends on the knowledge and experience of health professionals. In this respect, an automated method to be developed for the ADHD classification model is of great importance for clinicians. In this study, the effect of data augmentation on ADHD classification model with deep learning was investigated. For this purpose, magnetic resonance images were taken from NPIstanbul NeuroPsychiatry Hospital and ADHD-200 database. Since the images were not sufficient in terms of training, data augmentation methods were applied and by convolutional neural network (CNN) architecture, these data were classified and tried to reveal the diagnosis of the disease independently from the non-objective experiences of the health professionals.