Theses supervised by Prof. Dr. Ömer Nezih Gerek
12 theses · Anadolu University, Eskişehir Technical Üniversity
Markerless augmented reality applications
Augmented Reality (AR) is the process of adding virtual computer generated information on view of a physical real-world environment in real-time. AR, which allows adding 3-dimensional models, animations, videos and various virtual materials which are generated in computer environment to be positioned on desired targets in the real world, allows for enabling and enriching the real world more functional and beneficial. In this thesis explains how AR technologies work, expresses and explains fields of application and gives a summary from history to present time and also presents a real-life demonstration of Mobile device-based AR. Thesis presents contributions for design and development of a fully working Markerless Augmented Reality Application that works on Android Mobile devices. For this purpose, were studied on several computer vision techniques such feature extraction-detection, SURF (Speeded Up Robust Features), Markerless tracking (Natural feature tracking) by using OpenCV Computer Vision library. As well as, were done 3D modelling, texturing, shading etc. by using OpenGL (Open Graphics Library) and also creating Wavefront OBJ 3D model files via Blender 3D. Finally in this thesis, several applications are presented that make use of these methods which is mentioned above. Firstly camera calibration, pose tracking and marker based augmented reality applications were worked for understanding augmented reality idea, then a markerless augmented reality application were developed on android mobile device by using Vuforia SDK for target tracking part of algorithm. Keywords: Augmented Reality, Computer Vision, Mobile Devices, Android Software Interface, SURF
Nonlinear data modeling methods for multidimensional signal analysis
In this dissertation, various novel stochastic and deterministic nonlinear data models are proposed for the analysis of the discrete signals that are defined in multidimensional spaces. The purpose of choosing the multidimensional space is to analyze the data according to multiparameters at the same time. The reason of analysing nonlinear methods is their compatibility with chaotic and nonlinear behaviour of the data which are measured for natural events and in this thesis natural events are taken as case study. The comparison of stochastic and deterministic methods gives the opportunity to choose the most suitable model for the data of handled problem. As stochastic models, one and multidimensional versions of Mycielski method and different versions of Markov Chain Models are proposed. As determinsitic models, multidimensional polynoms, multidimensional splines, multidimensional Empirical Mode Decomposition and Wavelets are chosen. In addition a Markovian error tuning model is designed as an infrastructure to test these models, which is inspired from time varying and time invariant versions of the Hidden Markov Model. These comparative works try to reveals the phenomenon underlies the natural events as wind speed, solar radiation, temperature which are taken as case study in this work.
Powerline detection for aircraft flight safety using image processing
In this thesis, several active methods have been proposed for the safety of air vehicles, which prevent air vehicles from colliding with electrical wires. In this context, Discrete Cosine Transform, Linear Binary Pattern, Histogram of Gradient and Convolutional Neural Network methods were used. The methods were tested on image databases, which consist of real life aerial images that were captured and curated by ourselves. The recognition performances of such scenes with these methods were compared. The effects of various pre-processing / pre-training on the classification and feature extraction steps were also examined.
Sınırlı bakış durumunda ara-imge sentezleme
Bu çalışma kamera barındıran nesnelerin fonksiyonlarını daha da geliştirmek için başlatılmıştır. Görüntüleme cihazlarının sınırlı bakış açılarından aldıkları görüntülerden daha fazla bilgi edinmek için yöntemler uygulanmıştır. Farklı iki bakış açısından edinilen resimlerle diğer ara görüntülerin oluşturulması hedeflenmiştir. Farklı iki görüntü arasında matematiksel bir bağ kurabilmek amacıyla işaretleyici kareler kullanılmıştır. Karelerin görüntüler içerisinde otomatik bulunması, kameraların pozisyonlarının hesaplanabilmesini sağlamıştır. Ara görüntülerin sentezlenebilmesi için resimler üzerinde anahtar noktalar bulunmuştur. Bu noktaların bulunması için AKAZE, SURF, BRISK, ORB gibi farklı yöntemler incelenmiş ve karşılaştırılmıştır. Bulunan noktaların farklı iki resim arasındaki bağlarının oluşturulması amacıyla Brute Force ve FLANN tabanlı eşleyici algoritmalar kullanılmıştır. Bu iki algoritmanın başarımları konusunda inceleme ve değerlendirme ilgili bölümlerde bulunabilir. Ara görüntülerin hesaplanması için iki resimden elde edilecek derinlik algısına ihtiyaç vardır. Bu hesaplama için epipolar geometri kuralları kullanılmıştır. Bilgi edinilemeyen görüntü alanları için bilinen noktaların üç boyutlu interpolasyon metotları ile gerçekteki nesnelere yaklaşım sağlanmaktadır. Böylece fotoğraflardaki nesnelerin üç boyutlu yapıları mümkün olduğunca tekrar oluşturularak farklı bakış görüntüleri oluşturulabilmiştir.
Improvement of deep neural networks by intelligent model initialization
Developments in deep learning leveraging the recent abundance of parallel computational power and visual data have resulted in significant advances in visual recognition performance. Among the methods that have emerged from this work, deep convolutional neural networks trained with a large amount of data in a supervised manner has been able to consistently deliver state of the art performance in various visual tasks. However, the amount of data this method requires limits its usefulness in real-world applications. In this study, we investigate methods to relieve deep convolutional neural networks from extreme data dependency. First, we show that transferring representations from ImageNet pretraining reduces overfitting even if the target data distribution is significantly different. In addition, we propose two approaches to stochastically generate training data using analytically designed models. The first approach is to generate entirely synthetic training data based on Gestalt principles, which is suitable when the target pattern to be recognized is low-level. Alternatively, if the target pattern to be recognized is high-level, training data is derived from an existing dataset such as ImageNet. By utilizing analytically designed elements, these two approaches inject knowledge to the model and reduce data-dependency.
Path guidance and obstacle detection techniques using monocular vision systems for automated guided vehicles (AGVS)
There are many problems with object detection for autonomous robots and automated guided vehicles (AGVs) used in factories. The main problems consist of issues such as routing and obstacle recognition. For the solution of these, laser, sight, vision based sensors have been practically used. In this study, the methods used through vision-based sensors, such as camera, are in the field of interest. These methods consist of two categories: traditional or machine learning based. Recently, the most popular of these techniques are deep learning-based techniques, which is a sub-field of machine learning. Deep learning techniques such as YOLO, SSD, R-CNN have left many other techniques out of use with their performance. However, these techniques may occasionally experience performance degradation especially during video processing. Better results can be obtained by adding a novel Sequential Decision Theory layer on top of the available deep learning techniques.
Ambulatory monitoring of ECG signals for arrhythmia detection using deep neural networks
Reliable arrhythmia detection plays a pivotal role in preventing adverse cardiac events, yet the earliest warning signs often manifest as subtle, patient‑specific perturbations in the electrocardiogram~(ECG). This study introduces a multiscale wavelet framework that converts modified lead‑II segments from the MIT–BIH Arrhythmia Database into time–frequency scalograms via the Continuous Wavelet Transform. Real and complex Morlet, Mexican Hat, and first‑order Gaussian wavelets are evaluated over low, mid, high and full frequency bands. The resulting images are processed under strict inter‑patient splits by state‑of‑the‑art backbones—ResNet‑50, DenseNet‑121, EfficientNet‑B0 and ViT‑B/16. In the five‑class (non‑AAMI) setting, the best hybrid configuration attains {83.0\,\%} overall accuracy while boosting minority‑class sensitivity by nearly 20 percentage points. Multiscale augmentation with Mexican Hat and Morlet scalograms elevates DenseNet‑121 accuracy from \(\sim60\,\%\) to {80.6\,\%}. Under the four‑class AAMI taxonomy Normal (N), Supraventricular (S), Ventricular (V), and Fusion (F), the same strategy—combined with a class‑weighted loss—yields {88–90\,\%} accuracy when models are trained on one patient subset and evaluated on the other, demonstrating robust generalization. These findings indicate that wavelet‑based time–frequency representations, when paired with modern deep networks, deliver tangible improvements for ambulatory ECG monitoring. Future work will address real‑time wearable deployment, adaptive scale selection, transfer learning from larger ECG corpora and ultra‑low‑latency inference.
Hareket dengelemeli ve lifting tabanlı video kodlayıcı tasarımı
Bu tezde, hareket dengeleme yöntemini kullanan yükseltme tabanlı hibrit bir kodlayıcı tasarlanarak görüntü kalitesinden fark edilir düzeyde ödün vermeden video sinyallerinin sıkıştırma oranının arttırılması üzerine çalışılmıştır. Bu amaçla video çerçeveleri zamansal doğrultuda gruplara ayrılmış ve zamansal yükseltme tabanlı ayrıştırma işlemi blok tabanlı hareket dengeleme ile birlikte kullanılarak çerçeveler öngörülmeye çalışılmıştır. Öngörü için kenar-uyarlamalı yükseltme ve blok karşılaştırma yöntemleri farklı düzeylerde kullanılmıştır. Üzerinde öngörü işlemi yapılmayan çerçeveler ve öngörü çerçevelerinin orijinallerinden farkları nicemlenerek kodlanmıştır. Kodlamanın başarımını değerlendirmek için PSNR ölçümleri kullanılmıştır. Deneysel sonuçlar, blok karşılaştırma düzeyi arttıkça daha fazla sıkıştırma elde edilirken PSNR değerinin kabul edilebilir seviyelerde kaldığını göstermiştir. Buna göre 37,04 : 1 oranında sıkıştırılarak kodlanan videolar 29 dB sinyal gürültü seviyesinde geri çatılmıştır.
Tekil dönüşüm matrisleri üreten blok dalgacık süzgeç tasarımı
Bu tezde tekil dönüşüm matrisi üreten dalgacık süzgeç bankaları için çeşitli tasarım yöntemleri ve algoritmalar geliştirilmiştir. İlinti yapısı bilinen sinyaller için en büyük kodlama kazancını veren dönüşüm kodlama yöntemi Karhunen?Loéve dönüşümü olduğundan süzgeç bankalarının hesaplanmasında Karhunen?Loéve dönüşüm matrisi esas alınmış ve verilen bir sinyal için en iyi sıkıştırmayı veren süzgeç bankasının bulunması hedeflenmiştir. Bu amaçla hem nümerik, hem de analitik yöntemler üzerinde çalışılmış ve kullanılan yöntemlerin başarımları deney sonuçlarıyla birlikte sunulmuştur. Blok Dalgacık Dönüşümü Ters Çevrimi adı verilen ve kafes parametre belirleme yöntemine dayanan yeni bir yöntemle, hedeflenen dördün yansımalı süzgeç bankalarının katsayılarının hesaplanabileceği gösterilmiştir.
Resimlerdeki hareket bulanıklığına neden olan filtre matrislerinin yapay sinir ağları kullanılarak sınıflandırılması
When capturing a moving object with the camera, motion blur occurs in the picture depending on the speed and direction of movement. The motion blur can be classified in terms of filter matrices and the picture can be restored using it. The purpose of this project is finding the filter matrices that causes to the motion blur using artificial neural networks. Once the filter matrix or point spread function (PSF) found, the motion-blurred images can be sharpened using well-known algorithms such as Wiener, Lucy Richardson Deconvolution. A few datasets that have lots of training images are blurred with the parameters (blur angle and length) of specific motion blur kernels and they are trained using VGG, ResNet and a network model that is shared in an article. Then, the prediction of the filter matrices is evaluated using validation datasets for every trained networks. The prediction accuracies for the trained networks are compared to show which network type is best for the purpose of this thesis. Some types of artificial neural networks and their structures are investigated for their suitability in terms of finding the accurate PSF that causes motion blur. In addition, some modifications are applied to the network to increase the prediction accuracy. This improved the estimation of the correct filter matrices.
Öznitelik seçimi ve makine öğrenmesi ile yoksulluk seviye karakterizasyonu
Targeting poverty requires access to accurate, timely and reliable quantitative data on socio-economic characteristics of households. However, in many developing countries, collecting accurate, timely, and reliable data on household characteristics is expensive, time-consuming, and unreliable, often requiring long and detailed surveys. Reliable data on economic status remain scarce in developing countries, hampering efforts to study these outcomes and to design appropriate policy responses to improve household welfare. In such situations machine learning algorithms can be of a great help. However, these models are normally designed in the form of black boxes; if the model is trained on a certain known data and predicted on unseen data, it doesn't give any information about the features that discriminate between classes. In other words, it is very tough to extract the features indicating that someone falls under specific category of poverty. Moreover, in poverty identification, measurement or classification, it is crucial to know how such features contribute to each class of poverty. Therefore, we designed an approach that extracts a subset of features that best characterize each poverty class, examines how this subset affect the chosen class and finally employ ensemble models to best classify between these classes. Through this approach we look at poverty from a multidimensional perspective contrary to a single dimension perspective defined as living on consumption expenditure of less than a predefined income threshold. The application and usefulness of our proposed framework is tested on a Costa Rican dataset collected from Kaggle website and provided by Inter-American Development Bank. Keywords: Poverty Characterization, Poverty Measurement, Poverty Identification, Multidimensional Poverty, Feature Extraction, Machine Learning.
Yüzey doku örüntüsü sınıflandırma amaçlı görüntü işleme
Conventional pattern classification systems have mostly employed binary classification methods where training features are extracted from multiple classes. This approach faces a challenge when it comes to systems which deal with imbalanced distribution of class samples, a typical characteristic of defect detection systems where there also exists a wide spectrum of possible defects. Using binary classifiers in such a scenario is bound to introduce uncertainties with respect to classifier performance as defects which had not been used in the training stage are encountered. One class classifiers have been proposed to overcome this challenge by using only normal samples to train the classifier. This thesis provides a comprehensive analysis of one class (i.e. unitary) classification to provide an empirical evaluation of the effects of defect spectrum in the feature space. Different unitary classifiers were compared to common binary classifiers and experimental results showed significant instability in the performance of the binary classifiers when classes occupy different regions in the feature space relative to the training classes. The performance of unitary classifiers was stable in all defect type scenarios.