Ing. Miroslav Čepek, Ph.D.

Ing. Miroslav Čepek, Ph.D.

Ing. Zdeněk Buk, Ph.D.

Department of Applied Mathematics

AWS Deepracer is a popular platform for developing autonomous racing cars using reinforcement learning. In this thesis, the aim is to develop a model for AWS Deepracer that can navigate tracks it has not seen before. The method, that was used to achieve such a model involved using a combination of techniques including data augmentation, and hyperparameter tuning. The model was trained on a set of tracks that were not included in the evaluation dataset and evaluated its performance on a separate set of tracks. The outcome of this thesis are two AWS Deepracer models, one slower but more careful one, and one faster but less accurate one. Both of these models, however, are able to run fairly accurately on a wide variety of tracks, including ones, unseen in training. The outcomes of this thesis allow future AWS Deepracer developers that might want to build a general model, to start with some insights into the process or use the already trained models from this thesis as a base for their own models.

Thesis on DSpace

Ing. Miroslav Čepek, Ph.D.

doc. Ing. Kamil Dedecius, Ph.D.

Department of Applied Mathematics

The work deals with the automatic recognition of vehicle number plates. The result of this work is a program that is able to recognize text on a single-line number plate with dark text on a light background. The output of the program is an image marked with all detected number plates and with recognized text. The program is written in Python and uses YOLO version 4 implemented using TensorFlow for number plate localization and Tesseract for character recognition. By using the OpenCV library, my implementation improved the results initially, using only Tesseract OCR and YOLO. The improvement in the correct recognitions was more than ten times better. Correct number plate recognition reached up to 31.5%. I have also found more applicable use-cases where the implementation proved successful.

Thesis on DSpace

Ing. Miroslav Čepek, Ph.D.

Mgr. Vojtěch Rybář

Department of Applied Mathematics

This thesis is dedicated to the analysis of graph neural network methods for the nodes and graph classification. Explores current libraries for working with graph neural networks such as StellarGraph, PyTorch Geometric and DGL. The graph algorithms Graph Convolutional Networks, GraphSAGE and Graph Attention Networks are tested and compared on selected datasets from the Open Graph Benchmark. The achieved results are compared with the state of the art results.

Thesis on DSpace

Ing. Miroslav Čepek, Ph.D.

Ing. Luděk Kopáček, Ph.D.

Department of Applied Mathematics

This bachelor thesis focuses on processing product descriptions to extract product compatibility information. The solution uses known machine learning models from the natural language processing field and its subtasks named entity recognition, relationship extraction, and question answering. A suitable dataset must contain annotations specifying the named entities and their relationships. After creating the dataset, there are applied selected machine learning models. I extracted product compatibility information with a 62.30% score, using just named entity recognition and relationship extraction. We decided to skip the question answering task because it would be out of scope for this bachelor thesis. This thesis brings a solution for leveraging pre-trained models to analyse product descriptions and extract their compatibility. In summary, there are described the possibilities for further research. As an appendix, there is a detailed description of the configuration file used for our model.

Thesis on DSpace

Ing. Miroslav Čepek, Ph.D.

doc. Ing. Ivan Šimeček, Ph.D.

Department of Applied Mathematics

Due to various reasons -- like privacy or security concerns -- it's not always possible to work directly with an original graph, such as bank transactions or social network interaction graph. These graphs are necessary for ML projects such as bank security systems for the detection of abnormal transactions, social network recommendation systems, and many other similar projects. This thesis aims to investigate the current existing graph generation techniques and assess how much synthetic graphs statistically correspond to the properties of the original ones. It also evaluates the feasibility of using structural embedding models for classification problems using synthetic graphs.

Thesis on DSpace

Ing. Miroslav Čepek, Ph.D.

Ing. Magda Friedjungová, Ph.D.

Department of Applied Mathematics

This thesis processes the problematics of tabular data generation while preserving dependencies. The main aim of the thesis is to create a generator, which can generate data with the same statistical properties as original data. In addition, the model must be able to generate both numerical and categorical data. The present thesis contains a detailed implementation of two generative models, namely Variational autoencoder and Generative adversarial networks. The experimental part compares these models with chosen generators from the Synthetic data vault library. Models outputs are evaluated based on results from classification and regression tasks. Although Generative adversarial networks repeatedly scored the best results, it is impossible to determine a better model. Both generators proved that they can generate tabular data while preserving desired features.

Thesis on DSpace

Ing. Miroslav Čepek, Ph.D.

Ing. Zdeněk Buk, Ph.D.

Department of Applied Mathematics

Within this thesis, a comprehensive literature survey of various autonomous driving methodologies and machine learning model architectures has been conducted, with a particular focus on object avoidance. The thesis goes on to explore the capabilities of the AWS DeepRacer autonomous racecar platform. This platform is utilized to investigate the feasibility of training end-to-end self-driving models focused on object avoidance using reinforcement learning. Two self-driving architectures were compared, namely a three-layer convolutional neural network and a five-layer convolutional neural network architecture. Furthermore, the impact of sensor choice on the autonomous object avoidance task was compared. Experiments in the simulated environment showed, that the three-layer convolutional neural network architecture, equipped with a stereo camera and LiDAR sensors performed the best. The model was subsequently deployed to the DeepRacer vehicle and demonstrated in the real world. The thesis successfully demonstrated the feasibility of training end-to-end autonomous models using the AWS DeepRacer platform and simulated environment.

Thesis on DSpace

Ing. Miroslav Čepek, Ph.D.

Rodrigo Augusto da Silva Alves, Ph.D.

Department of Applied Mathematics

The growing popularity and importance of machine learning adoption across industries have led to a gradual enrichment of DevOps principles with data- and model-related concepts, forming a paradigm known as MLOps. This diploma thesis explores its importance and describes the main principles and phases involved. We perform a summary of MLOps tools and their features, which we use to select the appropriate tools for use in our academic setting, and design a proof of concept solution that can be used as a basis for further research and incorporation of machine learning operations to simplify the model development process for students and researchers at our university.

Thesis on DSpace

Ing. Miroslav Čepek, Ph.D.

Ing. Magda Friedjungová, Ph.D.

Department of Applied Mathematics

Digital image watermarking is a widely used technique for protecting intellectual property or authenticating digital media, but it can negatively impact image quality and usability. This motivates the need for removing watermarks from images, and deep learning presents a potential solution. This thesis develops a deep learning method for watermark removal, including a survey of existing techniques and the proposal of a novel architecture. The method's performance is evaluated in terms of watermark detection accuracy and image reconstruction quality.

Thesis on DSpace