|
|
Optical Character Recognition
Pokorný, Pavel ; Juránek, Roman (referee) ; Mlích, Jozef (advisor)
This paper presents some of the methods for locating and recognizing text in an image document. It describes feature extraction issues and commonly used machine learning algorithms. In the latest part, there is description of design and implementation of application for printed text recognition.
|
|
|
Attributes Calculation for Prediction of Mutation Effect on Protein Function
Šinkora, Jan ; Filák, Jakub (referee) ; Jaša, Petr (advisor)
This thesis deals with issues of bioinformatics, machine learning, algorithms and data structures. The thesis is based on existing applications, Caver and Deleterious, developed by students from the Faculty of Informatics, Masaryk University and the Faculty of Information Technology, Brno University of Technology. The Deleterious framework calculates protein attributes that are important for the prediction of the effect of protein mutations on its function. Caver is a tool that finds tunnels in the 3-dimensional model of a protein. The goal of the thesis is to extend these applications by adding more attributes to the prediction process that could lead to improved prediction. The added attributes are related to detection and measurement of protein pockets.
|
|
|
Reduction of Computation Cost in libSVM Using String Kernel Functions
Kubernát, Tomáš ; Sehnalová, Pavla (referee) ; Michlovský, Zbyněk (advisor)
The goal of this thesis was to implement four string functions into the library libSVM . Then apply series of tests with variable parameters values affecting the individual string functions using the library and string functions. Using the results of experiments the speed and success of clasification of my implementation of string functions in library libSVM was compared with the implementation of string functions in program kernels . In this thesis there are also described procedures of all tests along with measured data and their graphical representation.
|
|
|
Evolutionary Design of Neural Networks
Beluský, Tomáš ; Vašíček, Zdeněk (referee) ; Minařík, Miloš (advisor)
The work deals with the development of the genetic algorithm, which designs the structure and learning of the neural networks. The fitness function also includes the number of hidden neurons, and thus we obtain the most optimal structure, which is reachable. The own versions of the operators are presented, which manage the entire process of evolution. The result of the work is a library for evolutionary design of neural networks. Moreover, graphical interface for setting parameters and displaying the results was created. In the experimental part the design is compared with other systems and algorithms. Finally, results are reviewed and the process for the following development of the system is outlined.
|
|
|
Machine-Learning Methods in Natural Language Processing
Vantuch, Marek ; Mrnuštík, Michal (referee) ; Otrusina, Lubomír (advisor)
Firstly, basic rules of tagging of the Czech language are described as well as problems connected to this field. Thereafter the focus of the thesis is put on the success rate of testing on the Czech corpus and at the same time trying to find the most suitable parameter values for using the features. After reaching a reasonable compromise between duration and accuracy, the value is then attempted to be improved using analysis of separate features and their eventual omission.
|
|
|
Building Model Generator for Open Street Maps
Galacz, Roman ; Poulíček, Zbyněk (referee) ; Polok, Lukáš (advisor)
This work concerns obtaining data from the maps provided by the project OpenStreetMap. The data are converted from the format of geographical latitude and longitude to the Cartesian coordinate system. This work also concerns building type recognition in build-up area which are situated on the downloaded map. Part of the work is a demonstration application which is able to model 3D geometry of the buildings, based on the results of the recognition algorithms and also creates a terrain in which these buildings are situated. Generated model is displayed using the OpenGL graphics library.
|
|
|
Simple Character Recognition
Hamrský, Jan ; Svoboda, Pavel (referee) ; Polok, Lukáš (advisor)
This work deals with the process of text location and recognition in an image document. It discusses the matter of feature extraction and its usage in machine learning. Portion of this work is devoted to design and implementation of application for simple character recognition of machine printed text.
|
|
|
Annotation of Network Traffic
Holakovský, Jan ; Novotňák, Jiří (referee) ; Žádník, Martin (advisor)
This thesis focuses on methods for network traffic classification. Furthermore the thesis analyzes the advantages and limitations of theese approaches. This work coveres development of a new tool for manual network traffic inspection and classification which uses a combination of selected classification approaches. At the end of this thesis results of conducted experiments are presented and some possible future improvements are proposed.
|
|
|
Optimization of Heuristic Analysis of Executable Files
Wiglasz, Michal ; Křoustek, Jakub (referee) ; Hruška, Tomáš (advisor)
This BSc Thesis was performed during a study stay at the Universita della Svizzera italiana, Swiss. This thesis describes the implementation of a classification tool for detection of unknown malware based on their behaviour which could replace current solution, based on manually chosen attributes'scores and a threshold. The database used for training and testing was provided by AVG Technologies company, which specializes in antivirus and security systems. Five different classifiers were compared in order to find the best one for implementation: Naive Bayes, a decision tree, RandomForrest, a neural net and a support vector machine. After series of experiments, the Naive Bayes classifier was selected. The implemented application covers all necessary steps: attribute extraction, training, estimation of the performance and classification of unknown samples. Because the company is willing to tolerate false positive rate of only 1% or less, the accuracy of the implemented classifier is only 61.7%, which is less than 1% better than the currently used approach. However it provides automation of the learning process and allows quick re-training (in average around 12 seconds for 90 thousand training samples).
|
|
| |
guest ::
