Theses

doc. RNDr. Dušan Knop, Ph.D.

Ing. Daniel Vašata, Ph.D.

Department of Applied Mathematics

The maps of elections framework is a methodology for visualizing and analyzing election datasets. So far, this framework was restricted to collections of election datasets that have equal numbers of candidates. In this thesis, we extend this framework to be able to deal with datasets containing different numbers of candidates. We do this in two ways: we extend an already existing positionwise distance, and we also develop the idea and implementation of a feature distance. We test the new distances on synthetic data generated using the Mapel package.

Thesis on DSpace

Ing. Jan Čech, Ph.D.

Ing. Magda Friedjungová, Ph.D.

Department of Applied Mathematics

This thesis proposes a method for pose and expression transfer between face images. Given a source and target face portrait, the designed network pro- duces an output image where the pose and expression from the source face image are transferred onto the target identity. The architecture consists of two encoders and a mapping network that maps the two inputs into the latent space of StyleGAN2, which generates a high-quality image. The training is self-supervised without the need for labeled data. Our method achieves close to real-time image generation while also enabling the synthesis of random identities with independently controllable pose and expression.

Thesis on DSpace

Ing. Karel Klouda, Ph.D.

Ing. Magda Friedjungová, Ph.D.

Department of Applied Mathematics

This work is devoted to automatic metrical analysis of Czech syllabotonic verse metrically tagged inside a large poetic corpus - the Corpus of Czech Verse. First, it reimplements the existing data-driven approach used by a program called KVĚTA. Later, it models the problem as a sequence tagging task and solves it using machine learning. The BiLSTM-CRF model is used, representing the current state of the art for many sequence tagging tasks. Many different input configurations are tested. In all experiments, the inputted syllables or word tokens are represented by Word2Vec word embeddings trained on training data. The results are evaluated by computing three different accuracies of the predictions: syllable-level accuracy, line-level accuracy, and poem-level accuracy. It is shown that using BiLSTM-CRF represents a great success. With the best input configurations, it produces better results than the KVĚTA reimplementation, with predictions achieving 99.61% syllable accuracy, 98.86% line accuracy, and 90.40% poem accuracy. The most interesting finding is that the best results are obtained by inputting sequences representing whole poems instead of individual poem lines.

Thesis on DSpace

prof. Dr. Ing. Petr Kroha, CSc.

Department of Applied Mathematics

This thesis is concerned with the research of information retrieval models, state-of-the-art word embedding techniques and their possible use for multilingual retrieval. Modern approaches to multilingual retrieval that build on word embedding techniques usually work with some transformation that converts word representations from one language to another. The aim of this thesis is to investigate a model that does not work with this transformation and instead directly extracts dependencies between translations. A prototype web search engine should then be built on top of this model. The work has met all expectations and the resulting model was able to represent bilingual translations directly without using any transformation. This was achieved using parallel translated European Union documents, which were linked at paragraph level for both languages. The prototype search engine then operated based on the learned word representations derived from this model.

Thesis on DSpace

prof. RNDr. Pavel Surynek, Ph.D.

doc. Ing. Pavel Hrabák, Ph.D.

Department of Applied Mathematics

In this work I propose a hierarchical system of multi-agent coordination for simulating evacuations. I distinguish between two types of agents. Leaders communicate between themselves and escort their swarms to a safe location, while followers follow their leader. For this multi-agent path finding algorithms are used. I propose multiple models varying both in leaders' behavior towards their respective swarms and followers' behavior considering their attempts at individual evacuation. In the work I perform experiments, whose results will show how parameters of agents' behavior influence the success of an evacuation. Results of these experiments will point out advantages of communication between leaders, problems, that might occur during an evacuation and their connection to agents' inadequate behavior.

Thesis on DSpace

Mgr. Vojtěch Rybář

Ing. Daniel Vašata, Ph.D.

Department of Applied Mathematics

In this thesis we reviewed uses of machine learning in computational fluid dynamics. We then implemented a state-of-the-art graph neural network to simulate the flow around an airfoil in 2D. We train the model at lower speeds and angles of attack and then extrapolate to higher ones. We trained a model that extrapolates with a small precision error and remains stable on long rollouts.

Thesis on DSpace

doc. RNDr. Ing. Marcel Jiřina, Ph.D.

Ing. Tomáš Kalvoda, Ph.D.

Department of Applied Mathematics

The objective of this bachelor's thesis is to design a camera system for scanning tire treads of passing vehicles in a standardized form. After researching existing solutions for scanning tire treads a new camera system is proposed, which captures and unwraps the tire tread of vehicles stopped at a traffic light. A convolutional neural network is used to find the location of the tire in an image. The camera system was tested in real traffic at multiple traffic lights. 78 % of vehicle tires are captured correctly. The implemented algorithms correctly process the tire tread 89 % of time. This new camera system can be used to collect large amount of tire tread samples, which might otherwise be impractical and expensive to obtain. This data is useful for example in forensic analysis, to pair a tire tread with tracks left on the ground.

Thesis on DSpace

Ing. Jakub Novák

Ing. Miroslav Čepek, Ph.D.

Department of Applied Mathematics

The paper compares algorithms designed for image segmentation. Three algorithms belonging to the category of superpixels (felzenszwalb, SLIC and quickshift), two representatives of active contour models (snakes and level sets), random walker, region adjacency graphs and Otsu thresholding are compared. For this purpose, defects appearing in the industry are described. A more general categorization splitting defects into groups is made. Subsequently, two defects are selected from each category. A suitable combination of algorithm parameters is found on the first defect. Different preprocessing and color representation of the image are accounted for when the search is carried out. With the found parameters, a segmentation of the second defect is performed. Thus, the generalization capability and the suitability of the algorithm for the defects of the given category are tested. The success of the segmentation is measured by the IOU metric. The more successful algorithms achieved an average IOU measured over all images of a single defect of 90 %. In generalization test an average IOU of 53 % was achieved in some cases.

Thesis on DSpace

Ing. Magda Friedjungová, Ph.D.

Ing. Daniel Vašata, Ph.D.

Department of Applied Mathematics

Automated evaluation of road quality can be helpful to authorities and also road users who seek high-quality roads to maximize their driving pleasure. This thesis proposes a model which classifies road images into five qualitative categories based on overall appearance. We present a new manually annotated dataset, collected from Google Street View. The dataset classes were designed for motorcyclists, but they are also applicable to other road users. We experimented with Convolutions Neural Networks, involving custom architectures and pre-trained networks, such as MobileNet or DenseNet. Also, many experiments with preprocessing methods such as shadow removal or CLAHE. Our proposed classification model uses CLAHE and achieves 71% accuracy on a test set. A visual check showed the model is applicable for its designed purpose despite the modest accuracy since the image data are often controversial and hard to label even for humans.

Thesis on DSpace

prof. RNDr. Pavel Surynek, Ph.D.

Dr. techn. Ing. Jan Legerský

Department of Applied Mathematics

This bachelor thesis studies current exact approaches to multi-agent collective construction problem, with emphasis on three-dimensional structures, built by agents repositioning passive blocks on a grid under the condition of gravity. A generalization of current fastest exact model is proposed, using mixed integer linear programming, to accommodate varying durations of agent actions. An application of the proposed model is used in conjunction with a solver to exactly optimize construction plan of user entered structures. The result is visualized in Minecraft, using program built upon project Malmo API. A series of experiments is performed on several small instances to measure relative construction makespan reduction, relative to the instances with one-timestep actions. Results show significant makespan reduction at action durations used for visualization in Minecraft.

Thesis on DSpace

Ing. Jakub Novák

Ing. Magda Friedjungová, Ph.D.

Department of Applied Mathematics

This thesis focuses on the automated detection and classification of defects on glass rods. Previous solutions of similar problems are analysed and a new approach is proposed. Four camera systems using advanced optical components and lighting are designed. Also, analysed defect detection and classification methods are modified and extended. One of them using a model based on the modern Faster R-CNN architecture is able to detect 83 % of defects and to correctly classify 77 % of them.

Thesis on DSpace

Ing. Magda Friedjungová

Ing. Daniel Vašata, Ph.D.

Department of Applied Mathematics

Restoring damaged regions in image data is a relevant and difficult problem, which gets proportionally harder with the severity of the damage. In the last few years we have seen promising progress in tackling this issue using deep learning models. This thesis verifies and compares different approaches to handling the image inpainting problem. Since generative adversarial networks are one of the most inventive and promising architectures, a survey on current methods was performed and two selected methods were reimplemented. Our implementations are compared to other inpainting methods using classification models. The presented results reflect the influence of damage type and damage severity on the ability of each of the considered methods to successfully inpaint a damaged image.

Thesis on DSpace

Ing. Jakub Novák

Ing. Mgr. Ladislava Smítková Janků, Ph.D.

Department of Applied Mathematics

The thesis is about extraction of text information from photographs of identity documents. Proposed method uses convolutional neural network U-Net trained to create segmentation mask to find ID card in natural scene. Text is detected within known areas and is read using Tesseract library. The character recognition precision is 98,4 % and recall is 92,6 % on 48 tested photos of documents.

Thesis on DSpace

Ing. Magda Friedjungová

Ing. Daniel Vašata, Ph.D.

Department of Applied Mathematics

Most labelled real-world data is not uniformly distributed within classes, which can have a severe impact on the prediction quality of classification models. A general approach is to overcome this issue by modifying the original data to restore the balance of the classes. This thesis deals with balancing image datasets by data augmentation using generative adversarial neural networks. The primary focus is on generating images of underrepresented classes in imbalanced datasets, which is a process known as class balancing. The aim of this thesis is to analyse and compare data augmentation techniques including standard methods, generative adversarial networks and autoencoders. Evaluation is done using classifiers trained on the original, unbalanced and augmented datasets. The results achieved suggest how the performance of the methods proportionately deteriorates with increasing imbalance rate and diversity of datasets.

Thesis on DSpace

doc. RNDr. Pavel Surynek, Ph.D.

Department of Applied Mathematics

A car chase is a scenario that requires dynamic maneuvers, fast reactions, and strategic trajectory planning. It tests the limits of an autonomous driving system. We developed an autonomous driving system that can chase another vehicle using only images from a single RGB camera. At the core of the system is a novel dual-task convolutional neural network capable of simultaneously performing object detection as well as coarse semantic segmentation. The system has been firstly tested in CARLA simulation. We created a new challenging publicly available chasing dataset collected by manually driving the chased car. Using the dataset, we showed that the system benefits from using coarse semantic segmentation. The system that uses the semantic segmentation was able to chase the pursued car on average 10% longer than other versions of the system. Finally, we integrated the system into a subscale vehicle platform built on a high-speed RC car and demonstrated its capabilities by autonomously chasing another RC car.

Thesis on DSpace

Ing. Daniel Vašata, Ph.D.

Ing. Karel Klouda, Ph.D.

Department of Applied Mathematics

Recent research has shown that it is possible to design a model that learns to translate entirely from monolingual texts. Even though the translation quality still lags behind the state-of-the art models trained on texts translated by humans, this line of research opens new doors for low-resource language pairs. This thesis provides an overview of unsupervised techniques for machine translation applicable in low-resource conditions. We apply the most promising approaches and compare their performance on the Czech-German language pair. Since the proposed methods depend on vector representations of words in a cross-lingual space, we experiment with these representations to show how much language-neutral information they carry.

Thesis on DSpace

doc. RNDr. Pavel Surynek, Ph.D.

Ing. Daniel Vašata, Ph.D.

Department of Applied Mathematics

The thesis focuses on the usage of artificial neural networks in near-optimal solving of the (N^2-1)-puzzle. In the first part of the thesis, possible applications of artificial neural networks in solving the puzzle are analyzed, and it is found that the most effective way is to use them as a heuristic for state-space search algorithms. Later in the thesis, several heuristics based on deep artificial neural networks are trained, and they are evaluated on a set of benchmarks. When used together with the A* algorithm, the solutions obtained with the new heuristics are optimal most of the time, while the number of expanded nodes is significantly lower than that in comparable admissible and non-admissible heuristics.

Thesis on DSpace

MSc. Juan Pablo Maldonado Lopez, Ph.D.

Department of Applied Mathematics

Abstraction is an important tool for an intelligent agent. It can help the agent act in complex environments by selecting which details are important and which to ignore. In my thesis, I describe a novel abstraction algorithm called Online Partition Iteration, which is based on the theory of Markov Decision Process homomorphisms. The algorithm can find abstractions from a stream of collected experience in high-dimensional environments. I also introduce a technique for transferring the found abstractions between tasks that outperforms a deep Q-network baseline in the majority of my experiments. Finally, I prove the correctness of my abstraction algorithm.

Thesis on DSpace

Ing. Daniel Vašata, Ph.D.

Ing. Karel Klouda, Ph.D.

Department of Applied Mathematics

This thesis is focused on the possibilities of using natural language processing methods to analyze comments on the news portal. The main goal is to compare the ability of BERT, Doc2vec, and Doc2vec with pretrained word vectors from BERT to examine the relevance between the comments and the content of an article from a news portal. Another goal is to use the text vector representations to detect anomalies via the Local outlier factor method. It was found by experiments, that the best model for text representation is BERT and that the pretrained word vectors have no positive impact on results in comparison of Doc2vec without pretrained vectors. Moreover, the Local outlier factor can detect anomaly comments and users when using vectors from BERT in contrast to Doc2vec text representations which are not good enough for anomaly detection and therefore often returns incorrect results.

Thesis on DSpace

Ing. Tomáš Řehořek, Ph.D.

Ing. Kamil Dedecius, Ph.D.

Department of Applied Mathematics

Collaborative filtering is one of the most successful techniques used in recommender systems. The basic algorithms utilize history of interactions between users and items. However, recommenders deployed in production often have at least one more dimension of data available--timestamp of the interaction. These interaction circumstances are collectively referred to as the context. This thesis exploits the additional information in order to improve overall recommender accuracy. Several novel approaches to incorporating context into traditional collaborative filtering are proposed. An evaluation framework is designed and proposed algorithms are extensively evaluated with different parameters and contexts on multiple datasets. Results show that even mostly static datasets benefit from the proposed context-aware approach. About 5-35 % recall increase was observed in comparison with traditional collaborative filtering algorithms.

Thesis on DSpace

doc. Ing. Pavel Kordík, Ph.D.

MSc. Juan Pablo Maldonado Lopez, Ph.D.

Department of Applied Mathematics

Account sharing in online streaming services has a negative impact on recommendation systems and consequently on the quality of services provided to the users. This thesis aims to design a method capable of detecting such accounts. To address this issue, we designed and implemented algorithms based on collaborative filtering and sliding window method revealing shared accounts. The proposed algorithms allow detection of other user's activity with high accuracy, which can be filtered and recommendation system gets only relevant data for a given user. Furthermore shared accounts can be restricted since there is a subscription that users try to avoid.

Thesis on DSpace

Ing. Petr Máj

RNDr. Jiřina Scholtzová, Ph.D.

Department of Applied Mathematics

l-calculus is a fundamental concept in computer science and as such is taught at almost all universities with a computer science programme, including FIT CTU. But for many students, learning the l-calculus and understanding its significance and impact on programming languages is a challenging task. This thesis describes a l-calculus evaluator and its front-end designed to help stu- dents understand l-calculus by treating it more like a programming language and by effortless integration with existing course materials.

Thesis on DSpace

doc. RNDr. Pavel Surynek, Ph.D.

Ing. Daniel Dombek, Ph.D.

Department of Applied Mathematics

The subject of this bachelor thesis is focused on the automatic music transcription (AMT). The basic information related to this problematic, the state of the art in the solution of AMT and the problems tackled in this research domain are summarized in the theoretical part of this work. The approaches to the evaluation of results obtained by different AMT systems are also given. It has been found that unsatisfactory results are reached by AMT till now. Resulting music representation (MIDI file or sheet music) contains many errors that make it impossible to use this technology in practice. Therefore, in the practical part of the thesis, an analysis of the most common errors generated by the automatic transcription of a wave audio recording into a format MIDI were analyzed. Based on this analysis, a method for automatic removal of these errors and modification of acquired MIDI file is proposed. The goal of these modifications was to improve the success of the automatic music transcription system. The success of the proposed method was evaluated by F-measure and edit distance of the musical strings. The experiments presented in the final part of the thesis showed that the proposed method of adjustment increases the similarity of the resulting sheet music with the original and contributes significantly to the better readability of the resulting sheet music.

Thesis on DSpace

Ing. Tomáš Řehořek, Ph.D.

doc. Ing. Pavel Kordík, Ph.D.

Department of Applied Mathematics

This bachelor's thesis describes the recommendation system and two major approaches, Collaborative filtering and Content-based recommendation. The new hybrid approach, which combines these two methods, is proposed. This method increases recall of content-based recommendation by up to 216% and allows more precise recommendation for newly added items, which suffers from the cold-start problem. This designed and implemented approach uses machine learning methods such as embedding or artificial neural networks, which will also be briefly introduced along with a way of evaluating the quality of the recommendation.

Thesis on DSpace

Ing. Tomáš Řehořek, Ph.D.

doc. Ing. Pavel Kordík, Ph.D.

Department of Applied Mathematics

This thesis deals with the approximation of the user-based k-nearest neighbor algorithm using locality-sensitive hashing methods and the application of these methods in recommender systems. First the thesis describes the recommendation systems, the collaborative filtering, the k-nearest neighbors algorithms and the method of their approximation using locality-sensitive hashing methods. Based on the analysis, framework was designed and implemented to test various parameterizations of locality-sensitive hashing methods. The precision of nearest neighbors algorithm or the success rate of recommending (by using the recall rate depending on the catalog coverage) can be tested in the framework. Described methods are tested on two separed databases using the framework. From the tests of individual methods and parameterizations, the possibilities of their combinations are derived to achieve optimal models. During the testing of the optimal models I was able to achieve very satisfactory results. At around 3 % of the reference solution time, the reached success rate was between 97 % and 99 %. Finally, I discussed the results and different findings from the testing of the individual methods.

Thesis on DSpace

doc. RNDr. Ing. Marcel Jiřina, Ph.D.

doc. Ing. Pavel Kordík, Ph.D.

Department of Applied Mathematics

The thesis deals with the efficiency of the process of extracting knowledge from the data. It focuses on optimizing the order of pre-processing methods and their parameters for the specific classifier. An algorithm has been designed to create a plan to pre-process input data so that the learned classifier achieves the highest accuracy of classification. The genetic algorithm was used to optimize and find the plan. A Java application implementing this algorithm has been developed to find a pre-processing plan for any numerical dataset using the following data preprocessing methods: discretization, normalization, dimensionality reduction, remote values removal, class balancing and missing values imputation. The results were tested on several real datasets. The algorithm improves the classification accuracy by an average of 4-9 %. This is a tool that allows you to fully automate the pre-processing process. Eventually it can be used as a help tool for a knowledge expert to create a pre-processing plan.

Thesis on DSpace