|
|
Use of higher-order cumulants for ECG analysis
Maršánová, Lucie ; Janoušek, Oto (referee) ; Ronzhina, Marina (advisor)
This work deals with using higher order cumulants for analysis ECG. In the first part of work is described principle of electrocardiography, followed by matemathical derivation of higher order cumulants, description of their properties and their use in current practice. In the next part of paper is described caltulation higher order statistic of ECG beats in Matlab programming environment. In the practical part of work are tested predicted properties. Distinctive properties are minimalization of amplitude and time shift and Gaussian noise. This properties of higher order cumulants enable lesser variance of beats in one class and easier clasification ECG. Calculation of cumulants from real ECG beats of various groups is then realized. Classification based on the original ECG cycles and cumulants is performed using artificial neural network. Results of these classification approaches are then compared and discussed.
|
|
|
Heart beat representation for classification
Smíšek, Radovan ; Janoušek, Oto (referee) ; Ronzhina, Marina (advisor)
Selection of ECG segment plays a significant role in design of a heart beat classifier. The type of selected segments influences the classification not only in regard to the type and maximum number of recognized pathological groups but also in regard to the complexity of classification model, which consequently creates indirect demands on the memory of the computer technology used as well as on the time needed for the classification. The thesis is focused on the comparison of success rates of the ECG heart beat classifications in different input segments. The input segments used were QRST, RST, ST-T, QRS, and T. The ECG signal was obtained from isolated rabbit hearts and divided into individual types according to the T-wave amplitude and changes in the ST segment. The signal subsequently enters the artificial neural network where it is classified into predefined types. The network used had twenty-four neurons in the first layer and one neuron in the second layer. Efficiency of the classification is in the conclusion of this thesis.
|
|
|
Classification of heart beats using artificial neuronal networks
Doležalová, Radka ; Vítek, Martin (referee) ; Ronzhina, Marina (advisor)
This work deals with using of artificial neural networks (ANN) for ECG classification. The issue of ECG and ANN technique are described theoretically at first, the next section describes use of Matlab to design ANN and graphical user interface. ECG data (namely QRST segments from the orthogonal X- lead from seven phases of the experiment) obtained from experiments in isolated hearts of rabbits are used for learning and testing of the classifier. The result of this work is the software with GUI that allows user to set various parameters and structure of ANN. After learning phase, ANN realized in this work able to classify cardiac cycles according to their morphology into seven groups.
|
|
|
Automatic detection of K-complexes in sleep EEG signals
Pecníková, Michaela ; Ronzhina, Marina (referee) ; Kozumplík, Jiří (advisor)
This paper addresses the problem of detecting K-complexes in sleep EEG. The study of sleep has become very essential to diagnose the brain disorders and analysis of brain activities. Since Kcomplex can have a wide variety of shapes it is very difficult to detect the K-complexes manually. In this paper, I present an automatic method for K-complexes detection based wavelet transform,TKEO and method for classification using feedforward multilayer neural network designed in Matlab. Detection performance reached the value approx. from 52,9 to 83,6 %.
|
|
|
Automatic detection of ischemia in ECG
Noremberczyk, Adam ; Potočňák, Tomáš (referee) ; Ronzhina, Marina (advisor)
This thesis discusses the utilization of the artificial neural networks (ANN) for detection of coronary artery disease (CAD) in frequency area. The first part of this thesis is orientated towards the theoretical knowledge. Describes the issue of ECG pathological changes. ECQ are converted to frequency area. Described statistical methods and methods for automatic detection of CAD and MI. Explained the issue of the perceptron and ANN. The second deals with use of Neural Network Toolbox MATLAB®. This part focuses on counting and finding suitable parameters and making connection of band. At the end of the thesis UNS is used to detect ischemic parameters and the results are discussed. Average values for the best settings are 100% accuracy.
|
|
|
Main Text Extraction from Web Documents
Mrózek, Daniel ; Burget, Radek (referee) ; Bartík, Vladimír (advisor)
This thesis deals with the main text extraction from the web documents in HTML format. It describes some methods that are already used and their separation. The goal of the practical part is to propose an algorithm for main text detection in HTML pages using primarily text features in combination with position features. Block classification is solved by multilayer perceptron. It also describes implementation of the proposed algorithm, the testing procedure and presentation of the obtained results.
|
|
|
Application of Neural Networks for Human Face Localization
Žák, Jakub ; Štancl, Vít (referee) ; Švub, Miroslav (advisor)
This paper describes aplication of multi layered neural network for solving problem of detection human face in static picture. This Method has good generalizational capabilities in general and there is no need to assembly complex models of analyzed data. There is also mentioned posibility of using neural network with changed architecture in this work.
|
|
|
Melody Harmonization
Trnkóci, Andrej ; Jaroš, Jiří (referee) ; Fapšo, Michal (advisor)
Computer scientists have long been considering music as a particularly interesting art Indeed, the history of computer music is almost as long as the history of computer science. Programs to compose music, or to make music" at various levels of the composition process have been designed since the 50s. This bachelor's thesis surveys the main approaches in the field of automatic harmonization, i.e. the problem of producing musical arrangements (scores) from given melodies, and focuses on the most widely used techniques to do so. The main goal of this paper is the issue of design and implementation of a software system for an automatic music harmonization which should learn the rules of harmony from the database of midi file. In the paper. In this thesis I describe existing systems for harmonization and furthermore I focus mainly on principles of machine learning - theory and application of Artificial Neural Networks and their use for harmonization.
|
|
|
Automatic Segmentation of Documents Stored as Images
Jakub, Dušan ; Španěl, Michal (referee) ; Szőke, Igor (advisor)
This work deals with dividing the documents stored as images into three groups of segments - background, text and graphics. It introduces various solutions and the method using Gabor filters and artficial neural networks is described in detail. The selection of apropriate settings of the filters and training parameters of the network is discussed. Connected components searching is used for improving the results. A classifier writen in C++ and OpenCV library is part of the work. The designed procedure is applied for segmentation of scanned scientific papers, but also the results of segmentation of more complex documents (advertisements, presentation slides) are presented.
|
|
|
Object Detection in Images
Vaľko, Tomáš ; Motlíček, Petr (referee) ; Švub, Miroslav (advisor)
Object detection in images is quite popular topic for years. What stands for it are a lot of works from this area of computer science. This thesis is about object classification, specifically human faces, which are one of the most interesting objects for processing. For classification we use neural networks, learned on face database. We study what influence has size of face database and preprocessing of digital image on neural network learning. This project implements simple face detector and localizator. It summarizes more and less successful results and indicates possible ways of system development in the future.
|
guest ::
