|
|
Klasifikace příkazů z EMG pomocí neuronové sítě
Zauška, Ján ; Šůstek, Martin (oponent) ; Szőke, Igor (vedoucí práce)
Táto práca sa zaoberá klasifikáciou 15 príkazov (krátkych slov) z malej dátovej sady nahranej pomocou sEMG elektród umiestnených na tvári a krku rečníka. V nahrávkach sú rozlíšené dva typy reči - audible speech, čo je klasická reč, a silent speech, teda reč, pri ktorej je potlačené vydávanie zvuku. Práca popisuje spracovanie EMG signálu, extrakciu príznakov, návrh klasifikátoru a výsledky klasifikácie. Ako klasifikátor bola použitá vlastná architektúra konvolučnej neurónovej siete. V práci sa tiež nachádza mnoho experimentov porovnávajúcich presnosť klasifikácie silent a audible speech.
|
|
|
Fast Analysis of Borders in Image
Kolesár, Matej ; Španěl, Michal (oponent) ; Herout, Adam (vedoucí práce)
This thesis focuses on the problem of detecting edges in natural images while maintaining high performance per image. First, the existing approaches are analysed and from them the relevant information is extracted. This information is then used to design two architectures that use convolutional neural networks. One architecture is based on RCF and enriches the output, while the other is a combination of RCF and RCN. This combination provides better up-sampling and enriches the output even more. Evaluation was performed on the BSDS500 dataset and the best result was for achieved for the model that combined RCF and RCN with an ODS score of 0.675.
|
|
|
Robotic Tracking of a Person using Neural Networks
Zakarovský, Matúš ; Lázna, Tomáš (oponent) ; Žalud, Luděk (vedoucí práce)
The main goal of this thesis was to create a software solution based on a neural network to enable detection of a person and its subsequent following. This was achieved via completion of the points of the assignment. First, a hardware solution and used libraries and application programming interfaces were described as well as the robotic platform supplied by the Robotics and AI group of BUT Department of Control and Instrumentation upon which the robot was built on. Next, a research of various neural networks used for person detection was conducted. Four detectors were described in detail. Some of them were tested on either a PC or a NVIDIA Jetson Nano computer. Afterwards, a software solution consisting of five programs was created to achieve goals such as, detection of the person using ped-100 neural network, real-world position with reference to the robot estimation using monocular camera and robot control to successfully follow a target. The output of this thesis is a robotic platform able to detect and follow a person that can be used in a real-world applications.
|
|
|
Playing Gomoku with Neural Networks
Slávka, Michal ; Kolář, Martin (oponent) ; Hradiš, Michal (vedoucí práce)
This thesis explores the usage of AlphaZero algorithm for the game of Gomoku. AlphaZero is a reinforcement learning algorithm, which does not require any existing datasets and is able to improve only by using self-play. It uses a tree search for policy improvement, which is subsequently used for training. This approach was able to defeat the previous state of the art methods. Generating training data of high quality requires a lot of computationally expensive iterations, which makes them algorithm slow to train. Experiments show that the strength of the play is growing with each subsequent iteration, this might indicate that it still has room for improvement with more training and that it has not reached its full potential.
|
|
|
Detekce palných zbraní v obrazu
Debnár, Pavol ; Drahanský, Martin (oponent) ; Dvořák, Michal (vedoucí práce)
Táto práca sa zaoberá detekciou palných zbraní v obraze. V teoretickej časti je najprv definovaný pojem palná zbraň a potom sú uvedené typy palných zbraní. Nasleduje vysvetlenie obrazového šumu, ktorý môže do značnej miery ovplyvniť výsledok detekcie, a sú uvedené spôsoby, ako ho redukovať. Ďalej sú definované algoritmy obrazovej detekcie, ktoré pracujú na báze neurónových sietí - konvolučné neurónové siete (CNN) a Single Shot Multibox Detector (SSD). Vysvetlené sú aj klasické algoritmy pre detekciu v obraze ako HOG+SVM a SURF. V ďalšej časti sú uvedené použité knižnice a softvér. Nasleduje experimentálna časť, kde uvedieme navrhnutý postup riešenia a databázu. Pre detekciu boli použité algoritmy HOG+SVM, SURF a SSD. Následne sú uvedené výsledky testov na databáze a videu, čomu nasleduje zhrnutie a možnosti rozšírenia do budúcnosti.
|
|
|
Interconnection of Restricted Boltzmann machine method with statistical physics and its implementation in the processing of spectroscopic data
Vrábel, Jakub ; Hrdlička, Aleš (oponent) ; Pořízka, Pavel (vedoucí práce)
In this work, connections between statistical physics and machine learning are studied with emphasis on the most basic principles and their implications. Also, the general properties of spectroscopic data are revealed and used beneficially for improving automatized processing of the data. In the beginning, the partition function of a Boltzmann distribution is derived and used to study the Ising model utilizing the mean field theory approach. Later, the equivalence between the Ising model and the Hopfield network (machine learning model) is shown, along with an introduction for machine learning in general. At the end of a theoretical part, Restricted Boltzmann Machine (RBM) is obtained from the Hopfield network. Suitability of applying RBM to the processing of spectroscopic data is discussed and revealed by utilization of RBM to dimension reduction of the data. Results are compared to the standard tool (Principal Component Analysis), with discussing possible further improvements.
|
|
|
Software pro detekci a rozpoznání registrační značky vozidla
Masaryk, Adam ; Hradiš, Michal (oponent) ; Špaňhel, Jakub (vedoucí práce)
Cieľom tejto bakalárskej práce je navrhnúť a vyvinúť softvér, ktorý dokáže detegovať a rozpoznávať registračné značky z obrázkov. Softvér je rozdelený na 3 časti - detekcia značky, spracovanie výstupu detektora a rozpoznanie znakov na registračnej značke. Detekciu a rozpoznanie sme sa rozhodli implementovať pomocou moderných metód využitím konvolučných neurónových sietí.
|
|
|
Anonymizace videa
Mokrý, Martin ; Bartl, Vojtěch (oponent) ; Herout, Adam (vedoucí práce)
Cieľom tejto práce je navrhnutie a vytvorenie automatického systému pre anonymizáciu videí. Tento systém na svoju činnosť využíva rôzne detektory objektov v obraze a taktiež aj aktívne sledovanie takto zdetekovaných objektov. Na zdetekované objekty je následne aplikovaná úprava, ktorá zabezpečí dostatočnú mieru anonymizácie. Hlavným prínosom takéhoto systému je urýchlenie anonymizácie videí, ktoré bude následne možné zverejniť.
|
|
|
Exploiting Approximate Arithmetic Circuits in Neural Networks Inference
Matula, Tomáš ; Mrázek, Vojtěch (oponent) ; Češka, Milan (vedoucí práce)
This thesis is concerned with the utilization of approximate circuits in neural networks to provide energy savings. Various studies showing interesting results already exist, but most of them were very application specific or demonstrated on a small scale. To take this further, we created a platform by nontrivial modifications of robust open-source framework Tensorflow allowing us to simulate approximate computing on known state-of-the-art neural networks e.g. Inception or MobileNet. We focused only on replacement of most computationally expensive parts of convolutional neural networks, which are multiplication operations in convolution layers. We experimentally demonstrated and compared various setups and even that we proceeded without relearning, we were able to obtain promising results. For example, with zero accuracy loss on Inception v4 architecture, we gained almost 8% energy savings which could be valuable, especially in low-power devices or in large neural networks with enormous computational demands.
|
|
|
Automatic Speech Recognition System Continually Improving Based on Subtitled Speech Data
Kocour, Martin ; Veselý, Karel (oponent) ; Černocký, Jan (vedoucí práce)
Today's large vocabulary speech recognition systems are very accurate. However, tens or hundreds of hours of manually transcribed speech are needed in order to train such system. This kind of data is often unavailable, or they even do not exist for the desired language. A possible solution is to use commonly available but lower quality audiovisual data. This thesis addresses the methods of processing such data for semi-supervised training of acoustic models. Afterwards, it demonstrates how to continually improve already trained acoustic models by using these practically unlimited data. In this work is proposed a novel approach for selecting data based on similarity with the target domain.
|
host ::
RSS.