M.S. Theses

Permanent URI for this collection

https://digitalarchive.library.bogazici.edu.tr/handle/123456789/11645

Browse

Now showing 1 - 20 of 505

3D face registration using multiple average models
(Thesis (M.S.)-Bogazici University. Institute for Graduate Studies in Science and Engineering, 2008., 2008.) Alyüz, Neşe.; Akarun, Lale.
Three dimensional (3D) face recognition is a frequently used biometric method and its performance is substantially dependent on the accuracy of registration. In this work, we explore registration techniques. Registration aligns two faces and make a comparison possible between the two surfaces. In the literature, best results have been achieved by a one-to-all approach, where a test face is aligned to each gallery face separately. Unfortunately, the computational cost of this approach is high. To overcome the computational bottleneck, we examine registration based on an Average Face Model (AFM). We propose a better method for the construction of an AFM. To improve the registration, we propose to group faces and register with category-specific AFMs. We compare the groups formed by clustering in the face space with the groups based on morphology and gender. We see that gender and morphology classes exist, when faces are categorized with the clustering approach. As a result of registering via an AFM, it is possible to apply regular re-sampling on the depth values. With regular re-sampling, improvements in recognition performance and comparison time were obtained. As another factor causing diversity in the face space, we explore expression variations. To reduce the negative effect of expression in registration and recognition, we propose a region-based registration method. We divide the facial surface into several logical segments, and for each segment we create an Average Region Model (ARM). Registering via each ARM separately, we examine regional recognition performance. We see that even though some regions such as nose or eye area are less affected by expression variations, no single region is sufficient by itself and the use of all regions is beneficial in recognition. We experiment with several fusion techniques to combine results from individual regions and obtain performance increase.
3D human pose estimation from multi-view RGB images
(Thesis (M.S.) - Bogazici University. Institute for Graduate Studies in Science and Engineering, 2019., 2019.) Temiz, Hüseyin.; Akarun, Lale.; Gökberk, Berk.
Recovery of a 3D human pose from cameras has been the subject of intensive research in the last decade. Algorithms that can estimate the 3D pose from a single image have been developed. At the same time, many camera environments have an array of cameras. In this thesis, after aligning the poses obtained from single-view images using Procrustes Analysis, median ltering is utilized to eliminate outliers to nd nal reconstructed 3D body joint coordinates. Experiments performed on the CMU Panoptic, MPI INF 3DHP, and Human3.6M datasets demonstrate that the proposed system achieves accurate 3D body joint reconstructions. Additionally, we observe that camera selection is useful to decrease the system complexity while attaining the same level of reconstruction performance. We also derive that dynamic camera selection has a more signi cant impact on reconstruction accuracy as against static camera selection.
3D shape generation and manipulation
(Thesis (M.S.) - Bogazici University. Institute for Graduate Studies in Science and Engineering, 2022., 2022) Dirik, Alara.; Uğur, Emre.; Yanardağ, Pınar.
Computer graphics, 3D computer vision and robotics communities have pro duced multiple approaches to represent and generate 3D shapes, as well as a vast number of use cases. These use cases include, but are not limited to, data encoding and compression, shape completion and reconstruction from partial 3D views. How ever, controllable 3D shape generation and single-view reconstruction remain relatively unexplored topics that are tightly intertwined and can unlock new design approaches. In this work, we propose a unified 3D shape manipulation and single-view reconstruc tion framework that builds upon Deep Implicit Templates [1], a 3D generative model that can also generate correspondence heat maps for a set of 3D shapes belonging to the same category. For this purpose, we start by providing a comprehensive overview of 3D shape representations and related work, and then describe our framework and pro posed methods. Our framework uses ShapeNetV2 [2] as the core dataset and enables finding both unsupervised and supervised directions within Deep Implicit Templates. More specifically, we use PCA to find unsupervised directions within Deep Implicit Templates, which are shown to encode a variety of local and global changes across each shape category. In addition, we use the latent codes of encoded shapes and metadata of the ShapeNet dataset to train linear SVMs and perform supervised manipulation of 3D shapes. Finally, we propose a novel framework that leverages the intermediate latent spaces of Vision Transformer (ViT) [3] and a joint image-text representational model, CLIP [4], for fast and efficient Single View Reconstruction (SVR). More specifi cally, we propose a novel mapping network architecture that learns a mapping between the latent spaces ViT and CLIP, and DIT. Our results show that our method is both view-agnostic and enables high-quality and real-time SVR.
A BCPL translator using recursive descent technique
(Thesis (M.S.)- Bogazici University. Institute for Graduate Studies in Science and Engineering, 1984., 1984.) Kadifeli, Vasıl.; Balman, Tunç.
This thesis describes the implementation of a BCPL translator on a CDCCYBER 170/815 computer system using Recursive Descent Technique. The translator has been written in Pascal. The output of the translator is CDC COMPASS assembler instructions. A minimal run-time library of COMPASS routines have been prepared to provide for the interface of the BCPL language to the operating system.
A biometric authentication technique using spread spectrum audio watermarking
(Thesis (M.S.) - Bogazici University. Institute for Graduate Studies in Science and Engineering, 2014., 2014.) Can, Yekta Said.; Alagöz, Fatih.
Watermarking has become important in the last decade because of the copyright protection applications. Embedding information into an audio le is more di cult as compared to images, because human auditory system is more sensitive than human visual system. Therefore, the proposed watermarking algorithms for digital audio have been less than those for digital image and video. This thesis presents a biometric authentication scheme based on spread spectrum watermarking technique. We add a biometric authentication system to the Sipdroid open source VoIP program. Firstly, senders must register to the system with their unique biometric features. T.C Identity number, keystroke dynamics and voice are used as biometric features. After registration, these biometric features are used as watermarked material. Before embedding, the watermark is spread with the Direct Sequence Spread Spectrum (DSSS) technique. While talking, this watermark material is embedded to speech and sent to receiver using Frequency Hopping Spread Spectrum(FHSS) technique. The watermarked biometric data is constructed in the receiver's phone after conversation is nished. This method does not need the original audio carrier signal when extracting watermark because it is using the blind extraction. The experimental results demonstrate that the embedding technique is not only less audible but also more robust against the common signal processing attacks like low-pass lter, adding white Gaussian noise, shearing, and compression. In order for receiver to be able to login to the system, biometric features of the user should match with the watermarked biometric data.
A blockchain based group key agreement protocol (B-GKAP)
(Thesis (M.S.) - Bogazici University. Institute for Graduate Studies in Science and Engineering, 2020., 2020.) Taçyıldız, Yaşar Berkay.; Alagöz, Fatih.
Group key agreement protocols are crucial in the case of multiple parties agreeing on a common key without a centralized entity. However, the decentralized characteristic of these protocols causes performance challenges where parties need to communicate and verify other participants in the group. To overcome this issue, we propose a new approach to the group key agreement protocols by utilizing Hyperledger Fabric framework as a blockchain platform. To this end, we migrate the communication and verification overhead of the group key agreement participants to the blockchain network in our developed scheme. This paradigm allows a flexible group key agreement protocol that considers resource-constrained entities and trade-offs regarding distributed computation. According to our performance analysis, participants with low computing resources can efficiently utilize our protocol. In addition, the secret parameters of the participants are distributed among the isolated participants that constitute the blockchain network. Thus, the only way for the network participants to compute group keys is to collude maliciously. Furthermore, we have demonstrated that our protocol has the same security features as other comparable protocols in the literature.
A common subexpression elimination-based compression method for the constant matrix multipication
(Thesis (M.S.) - Bogazici University. Institute for Graduate Studies in Science and Engineering, 2022., 2022) Bilgili, Emre.; Yurdakul, Arda.
The execution time, resource and energy costs of deep learning applications become much more important as their popularity grows. The Constant Matrix Multi plication has been studied for a long time and takes place in deep learning applications. Reducing the computation cost of those applications is a highly active research topic. The weights are pruned or quantized while satisfying the desired accuracy requirement. The pruned matrices are compressed into one-dimensional arrays without data loss. Matrix multiplication is performed by processing those arrays without decompression. Processing one-dimensional arrays to perform matrix multiplication is deployed on vari ous hardware platforms that employ Central Processing Unit, Graphics Processor Unit and Field-Programmable Gate Array. The deployments can also be supported with common subexpression elimination methods to reduce the number of multiplications, additions and storage size. However, the state-of-the-art methods do not scale well for the large constant matrices as they reach hours for extracting common subexpressions in a 200 × 200 matrix. In this thesis, a random search-based common subexpression elimination method is constructed to reduce the run-time of the algorithm. The algo rithm produces an adder tree for a 1000 × 1000 matrix in a minute. The Compressed Sparse Row format is extended to build a one-dimensional compression notation for the proposed method. Simulations for a single-core embedded system show that the latency is reduced by 80% for a given 100×100 matrix compared to the state-of-the- art methods. The storage size of the sparse matrices is also reduced by more than half in the experiments compared to the Compressed Sparse Row format.
A comparative evaluation of machine learning algorithms for statistical downscaling of monthly mean temperature data over European region
(Thesis (M.S.) - Bogazici University. Institute for Graduate Studies in Science and Engineering, 2022., 2022) Eser, Günay.; Kurnaz, M. Levent.
Climate change is the most vital environmental change that has already started to affect many ecosystems. It is caused by greenhouse gas emissions which are increasing since the pre- industrial era, and populated areas become more vulnerable to disasters due to climate change. It has never been more crucial to model the climate effects on local regions. Organizations like Intergovernmental Panel on Climate Change (IPCC) use global climate models (GCMs) to project future changes in climate on a continental scale. Although these models are becoming more accurate, downscaling these models to smaller scales is an important task that is studied by climate scientists. The two main downscaling methods are dynamical and statistical downscaling. Statistical downscaling studies are more reachable and important to develop when compared to dynamical downscaling due to its lower costs. The use of machine learning algorithms in statistical downscaling is a new area. Studies that implement machine learning to make local scale projections of surface temperature are numbered. In this paper, four different machine learning algorithms were tested on downscaling of two different surface temperature datasets over a European region with different resolutions. The best performing algorithm was also tested augmenting elevation data. The results show that Gaussian process regression performs the best with MAE of 0.04 - 0.51 as compared to the other machine learning algorithms tested. In conclusion, machine learning algorithms such as Gaussian process regression can be a suitable approach when downscaling spatial monthly mean surface temperature data.
A comprehensive analysis of subword tokenizers for morphologically rich languages
(Thesis (M.S.) - Bogazici University. Institute for Graduate Studies in Science and Engineering, 2022., 2022) Erkaya, Erencan.; Güngör, Tunga.
Transformer language models have paved the way for outstanding achievements on a wide variety of natural language processing tasks. The first step in transformer models is dividing the input into tokens. Over the years, various tokenization ap proaches have emerged. These approaches have further evolved from character and word-level representations to subword-level representations. However, the impact of tokenization on models performance has not been thoroughly discussed, especially for morphologically rich languages. In this thesis, we comprehensively analyze subword tokenizers for Turkish, which is a highly inflected and morphologically rich language. We define various metrics to evaluate how well tokenizers encode Turkish morphol ogy. Also, we examine how the tokenizer parameters like vocabulary and corpus size change the characteristics of tokenizers. Additionally, we propose a new tokenizer for agglutinative and morphologically rich languages. We demonstrate that our tokenizer reduces overall perplexity and enables better generalization performance. Downstream task experiments show that morphology supervision in tokenization improves model performance.
A comprehensive analysis of using WordNet, part-of-speech tagging, and word sense disambiguation in text categorization
(Thesis (M.S.) - Bogazici University. Institute for Graduate Studies in Science and Engineering, 2012., 2012.) Çelik, Kerem.; Güngör, Tunga.
By the huge increase of data volume in the digital environment and the machine learning techniques, studies on automatic categorization of text documents is increased. Text categorization is simply assigning prede ned label to unseen documents by using some learning models. Traditional text categorization is based on statistical analysis of documents to represent the document with some vectors. And then, one of the machine learning techniques is used for categorization of documents.In addition to the traditional text categorization techniques, in this thesis, we group words by their part of speech tag and investigate the e ect of each part of speech individually and jointly in the classi cation accuracy. Furthermore, we incorporate semantic features such as synonyms, hypernyms, hyponyms, meronyms and topics into the documents by using WordNet. Thus we add meaning of terms. One of the problems faced in this study is that not all the semantic features really related to the document, in other words synsets generate ambiguity. To solve the problem we introduce a new method to eliminate the ambiguity. In this thesis the main objective is to investigate the contribution of semantic features. By incorporating semantic features we add meaning to the documents and thus the classi cation accuracy increased.
A content based microblogger recommendation model
(Thesis (M.S.) - Bogazici University. Institute for Graduate Studies in Science and Engineering, 2012., 2012.) Çelebi, Hüseyin Burak.; Üsküdarlı, Suzan.
Social networks are one of the most signi cant information sources on the Internet. People share information, their feelings, their opinions and interesting links. A microblogging system is a special kind of social network in which users post short but frequent update messages. Microbloggers subscribe (follow) to posts of others. However, nding relevant microbloggers to follow is a major problem, due to the massive quantity of users as well as the di culty of mentally aggregating fragmented short contributions. In this thesis, a content based recommendation model is proposed, which given a query recommends a set of ranked microbloggers. This model focuses on the content of posts as well as other characteristics of microbloggers to evaluate the relevance of microbloggers to the query. This thesis describes the model and a prototype implementation. Finally the outcome of a test with 41 users is discussed along with observations and recommendations for improved recommendations.
A cultural market model
(Thesis (M.S.)-Bogazici University. Institute for Graduate Studies in Science and Engineering, 2007., 2007.) Herdağdelen, Amaç.; Bingöl, Haluk.
A decentralized framework with dynamic and event-driven container orchestration at the edge
(Thesis (M.S.) - Bogazici University. Institute for Graduate Studies in Science and Engineering, 2022., 2022) Özyar, Umut Can.; Yurdakul, Ahmet.
Persistent advancements are being made at a rapid pace on enabling edge com puting for the Internet of Things technology to capitalize on. Vendors and developers are exploring new techniques to smooth out the process of navigating through the inherent heterogeneity of the edge networks. However, application delivery, resource allocation, fault tolerance, and security issues are yet to be fully solved while also pro viding a seamless experience for consumers. With virtualization and lightweight con tainer management platforms providing an abstraction layer, it is possible to deploy the same application on devices with different architectures and achieve uniformity. Towards a fully decentralized edge, the framework proposed in this thesis lays down the groundworks for dynamic container orchestration. It provides a blockchain based delivery platform for container applications with their updates and resource specifications through a registry on a distributed file system, namely InterPlanetary File System (IPFS). Then, enabled by the operating system virtualization, the framework handles resource allocation, container availability and scaling. A self-adaptive resource manager running on the metrics scraped from the host and the virtualization platform, i.e. Docker in our implementation, dynamically optimizes the resources allocated to each container. The framework ensures that variable workloads of a heterogeneous environment can co- exist on an edge device that is designed to be further extended to multiple devices. To achieve a truly distributed system, an event-driven architecture is built over a lightweight messaging protocol, MQTT, capitalizing on the asynchronous and distributed nature of the publish/subscribe pattern.
A defect prediction method for software versioning
(Thesis (M.S.)-Bogazici University. Institute for Graduate Studies in Science and Engineering, 2006., 2006.) Kastro, Yomi.; Bener, Ayşe B.
Software lifecycle is becoming more human-independent with the help of new methodologies and tools. Many of the research in this field focus on defect reduction, defect identification and defect prediction. Defect prediction is a relatively new research area using various methods from artificial intelligence to data mining. Currently, software engineering literature still does not have a complete defect prediction solution for new versions of a software product. In this research our aim is to propose a model for predicting the number of defects in a new version of a software product relative to the previous version by considering the changes. These changes might be introduced as a new feature or a change of algorithm or even as a form of a bug fix. Analyzing the types of changes in an objective and formal manner and considering the lines of code change, we aim to predict the new defects introduced into the new version. Using such a proposed model will benefit to a more focused testing phase which will decrease the overall effort and cost. Also, this method can help to determine the stability of a software version before publishing the product. The method also helps us to understand the individual effect of a feature, bug fix or change in terms of probability of a new defect introduction.
A drawing tool for protein interaction maps in Kohn notation
(Thesis (M.S.) - Bogazici University. Institute for Graduate Studies in Science and Engineering, 2011., 2011.) Edes, Mine.; Özturan, Can.
Cancer is one of the most lethal and common diseases of today. Causes of cancer have been studied widely as it is an essential part of cancer research. Cancer is mainly caused by malfunctioning of the tumor suppressor proteins. Generally, these proteins take part in protein-protein interaction networks. Hence, to understand the suppressor proteins, it is important to study these networks. Results from this study will have great impact on the cancer research, drug design and protein engineering. For that reason, these networks are represented in network diagrams called Protein Interaction Maps to visualize and understand them better. Today, there is still a need for the standard for visualization of the protein interactions. For this reason, Kohn and his group’s MIM (Molecular Interaction Map) notation is considered to be the answer to that need. Even though there are some tools for graphical visualization of protein interactions, there is no tool that can draw protein interactions with MIM notation with full support. Thus, in this study we aimed to design a tool that can draw with Kohn’s notation. We developed MIMTool; a drawing tool for manually drawing protein interaction maps in Kohn notation. Later, as one of the most important part of our study, we added a semi-automatic map drawing feature to the tool. This feature automatically draws the interactions between physical entities using Dijkstra’s shortest path algorithm. With MIMTool, it will be much faster to draw, update and exchange molecular interaction maps. Use of this tool will save time and decrease work load.
A dynamic saliency based method for video retargeting
(Thesis (M.S.) - Bogazici University. Institute for Graduate Studies in Science and Engineering, 2015., 2015.) Koçberber, Hatice Çiğdem.; Salah, Albert Ali.
With increased usage of smartphones, tablets and small displays to play multimedia content, video retargeting becomes an important tool for better user experience. In this thesis, we propose a novel content-based approach for video retargeting that relies on spatio-temporal saliency to estimate relevant information in videos. Our method preserves spatial saliency as well as temporal coherence. We also propose a spatio-temporal saliency algorithm designed for this application domain that combines spatial saliency with motion trajectories. We demonstrate the quality of the proposed approach through quantitative and qualitative evaluation, contrasting it with ve di erent video retargeting methods. Quantitative evaluation is done using generic image/video quality metrics, so that they can be applied on any video retargeting solution. We have extracted the correlation between the quantitative and qualitative evaluation, to propose a new metric that is a combination of the existing quantitative metrics. The proposed metric is proven to be the best approximation to the qualitative results, thus can be used as a benchmark to evaluate video retargeting methods.
A framework to improve user story sets through collaboration
(Thesis (M.S.) - Bogazici University. Institute for Graduate Studies in Science and Engineering, 2023., 2023) Köse, Salih Göktuğ.; Aydemir, Fatma Başak.
Agile methodologies have become increasingly popular in recent years. Due to its inherent nature, agile methodologies involve stakeholders with a wide range of expertise and require interaction between them, relying on collaboration and customer involvement. Hence, agile methodologies encourage collaboration between all team members so that more efficient and effective processes are maintained. Generating requirements can be challenging, as it requires the participation of multiple stakeholders who describe various aspects of the project and possess a shared understanding of essential concepts. One simple method for capturing requirements using natural language is through user stories, which document the agreed-upon properties of a project. Stakeholders try to strive for completeness while generating user stories, but the final user story set may still be flawed. To address this issue, we propose SCOUT: Supporting Completeness of User Story Sets, which employs a natural language processing pipeline to extract key concepts from user stories and construct a knowledge graph by connecting related terms. The knowledge graph and different heuristics are then utilized to enhance the quality and completeness of the user story sets by generating suggestions for the stakeholders. We perform a user study to evaluate SCOUT and demonstrate its performance in constructing user stories. The quantitative and qualitative results indicate that SCOUT significantly enhance the quality and completeness of the user story sets. Our contribution is threefold. First, we develop heuristics to suggest new concepts to include in user stories by considering both the individuals’ and other team members’ contributions. Second, we implement an open-source collaborative tool to support writing user stories and ensuring their quality. Third, we share the experimental setup and materials to evaluate the SCOUT.
A gan-based hybrid data augmentation framework on chest x-ray images and reports
(Thesis (M.S.) - Bogazici University. Institute for Graduate Studies in Science and Engineering, 2022., 2022) Özfidan, Hasan Berat.; Yanardağ, Pınar.; Özgür, Arzucan.
Classical data augmentation techniques are widely used by many image classifi cation applications in the absence of adequate training data. These data augmentation techniques consists of but not limited with reflection, random cropping, re-scaling exist ing images and transformations. These techniques are widely used in practice during training classifiers with extended versions of real-world datasets. Increasing dataset size with realistic synthetic data allows us to improve the classification accuracy by making use of additional realistic variety. With the great representational power of GANs, learning the distribution of real data with a consistent level of variety allows us to generate samples with nearly-unobserved discriminative features. In our ap proach we used the aforementioned generative capability of GANs by utilizing state of the art GAN augmentation framework titled as StyleGAN2-ADA. After the training SytleGAN2-ADA in class conditional setting, we extended the dataset with different numbers of additional generated samples in order to observe the correlation of accu racy and augmentation strength. We extended our approach by using StyleCLIP to experiment disentangled feature augmentations which is a novel approach in the field of GAN augmentation. To make use of StyleCLIP more efficiently, we fine-tuned CLIP with X-ray images and modified entities which are extracted from corresponding med ical reports. We used the DeepAUC framework which is proven to be efficient for multi- disease labelled X-ray classification tasks to test the performance of the GAN augmentation. In our approach, we observed that the classification accuracies were improved compared to without text-manipulated GAN augmented setting.
A general object tracker for locating objects in digital video
(Thesis (M.S.)-Bogazici University. Institute for Graduate Studies in Science and Engineering, 2008., 2008.) Fidaner, Işık Barış.; Akarun, Lale.
Tracking a human head in a complicated scene with changing object pose, illumination conditions, and many occluding objects, is the subject of this thesis. A general tracking algorithm is presented, which uses a combination of object color statistics and texture features with motion estimation. The object is defined by an ellipse window that is initially selected by the user. Color statistics are obtained by calculating object color histogram in the YCrCb space, with more resolution reserved for chroma components. In addition to the conventional discrete color histogram, a novel method, Uniform Fuzzy Color Histogram (UFCH) is proposed. The object texture is represented by lower frequency components of the object's discrete cosine transform (DCT), and local binary patterns (LBP). By using the tracker, performances of different features and their combinations are tested. The tracking procedure is based on constant velocity motion estimation by condensation particle filter, in which the sample set is obtained by the translation of the object window. Histogram comparison is based on Bhattacharyya coefficient, and DCT comparison is calculated by sum of squared differences (SSD). Similarity measures are joined by combining their independent likelihoods. As the combined tracker follows different features of the same object, it is an improvement over a tracker that makes use of only color statistics or texture information. The algorithm is tested and optimized on the specific application of embedding interactive object information to movies.
A hybrid BERT-GAN system for protein - protein interaction extraction from biomedical text
(Thesis (M.S.) - Bogazici University. Institute for Graduate Studies in Science and Engineering, 2021., 2021.) Basmacı, Mert.; Özgür, Arzucan.
Considering the rapid increase in the biomedical literature, manual extraction of information regarding Protein-Protein Interactions (PPIs) becomes an exhausting task. Therefore, there is a strong need for the development of automatic relation extraction techniques from scientific publications. In this study, we introduce a novel two-stage system to extract PPIs from biomedical text. Our approach contains two cascaded stages. In the first stage, we utilize a transformer-based model, BioBERT, to determine whether pairs of proteins appearing in a sentence interact with each other; therefore, we perform a binary relation extraction task. In the second stage, we adopt a Generative Adversarial Network (GAN) model that consists of two contesting neural networks to eliminate false-positive predictions of the first stage. We evaluate the performance of both stages separately on five benchmark PPI corpora: AIMed, BioInfer, HPRD50, IEPA, and LLL. Later on, we combine the five corpora into a single source to examine the system performance on a general PPI corpus. Finally, we apply our system to a case study for Host-Pathogen Interaction extraction from the COVID-19 literature. The experimental results show that our first stage achieves the state-of-the-art F1-score of 79.0% on the AIMed corpus and obtains comparable results to previous studies on the other four corpora. Moreover, our second stage results reveal that the GAN model improves the first stage results when our BioBERT model is trained on the combined corpus. Our case study results demonstrate that the proposed system can be useful as a real-world application.

Browse

Browsing M.S. Theses by Title

Results Per Page

Sort Options