|
|
Adversarial Augmentation for Robust Speech Separation
Pavlus, Ján ; Černocký, Jan (referee) ; Žmolíková, Kateřina (advisor)
Separace řečníků se zabývá separácí signálů jednotlivých řečníků z dané směsi vícero řečníků. Neuronové sítě trénované pro separaci řečníků fungují většinou dobře na uměle smíchaných nahrávkách, ovšem při použití směsí z reálného světa často selhávají. Pro zlepšení tohoto chování, je možné použít augmentaci trénovacích dat, jako je například přidání šumu. Nicméně tyto augmentace jsou limitovány tím, že musí být ručně navrhnuty. V této práci je použita modifikovaná verze modelu generativních adversarialních sítí (GAN), která může zlepšit tuto vlastnost tak, že generuje augmentace na základě míry zmatení separačního systému. Po každém kroku trénování generátoru a separátoru se systém separace řečníků stává více robustní. Takto navrhnutý model byl podroben experimentům. Během těchto experimentů byly různě nastavovány parametry GAN modelu, aby se nalezlo jejich nejlepší nastavení, které by vedlo ke správnému natrénování modelu, bez zkolabování do žádného módu. Během experimentů bylo takové nastavení nalezeno. Z takto natrénovaného modelu byl vybrán nejvíce robustní separátor a ten poté vyhodnocen. Výsledky hodnocení neukázaly zlepšení funkčnosti zrobustněného separačního systému vůči samému nezrobustněnému systému předtrénovanému na WSJ0-2mix datasetu, během testování na datasetu WHAM. Nicméně výsledky jiného hodnocení experimentů ukázaly, že separátor vybrán z trénování GAN modelu je značně zrobustněn oproti původnímu.
|
|
|
Generative Neural Networks for Handwritten Text
Ševčík, Pavel ; Dobeš, Petr (referee) ; Hradiš, Michal (advisor)
The aim of this study was to create a generative neural network for handwritten text lines. The model produces variable-sized images of handwritten text lines based on the expected style. The proposed method exceeds existing models in the image quality and can be used to generate both individual words and entire lines of handwritten text. It combines the use of the attention mechanism to extract the features for each character from the text query and their arranging on the line by inserting spaces between them. The new approach allows more granular control of the symbol positions on the line, which leads to smoother style interpolations. In contrast to the previous approach, the proposed method uses the Gaussian filter to spread the individual symbols features to the surrounding area. This approach also allows to train the model for symbols position predictions using the adversarial loss (GAN). In addition, annotations of symbol horizontal positions on the lines of the IAM dataset of handwritten text have been created.
|
|
|
Learning the Face Behind a Voice
Zubalík, Petr ; Mošner, Ladislav (referee) ; Plchot, Oldřich (advisor)
The main goal of this thesis is to design and implement a system that will be able to generate a face based on the speech of a given person. This problem is solved using a system composed of three convolutional neural network models. The first one is based on the ResNet architecture and is used to extract features from speech recordings. The second model is a fully convolutional neural network which converts the extracted features into the styles which form a base for the final facial image. These styles are then passed as an input to the StyleGAN generator, which creates the resulting face. The proposed system is implemented in the Python programming language using the PyTorch framework. The last chapter of the thesis discusses some of the most significant experiments performed to fine-tune and test the developed system.
|
|
|
Image Super-Resolution Using Deep Learning
Mojžiš, Tomáš ; Beran, Vítězslav (referee) ; Španěl, Michal (advisor)
The aim of this thesis is to create a deep neural net capable of super-resolution on images acquired by electron microscopes. The thesis consists of two parts - finding appropriate data and creating a dataset for the super-resolution task and designing a neural net architecture capable of solving the super-resolution task. Within the thesis, two datasets comprised of images acquired by electron microscopes were created. The datasets differ in the approach to data augmentation. They allow to train a neural network which fulfills the super-resolution task. To solve this task, two U-Net based and one GAN based architecture were trained. The resolution of images was upscaled by a factor of two and four. The best artificially upscaled images were created by neural network Real-ESRGAN. The values of metrics were not higher than the tested interpolation method, but the images seem more visually pleasing especially when they were upscaled four times. Thanks to this thesis, two datasets were created allowing to train other possible neural network architectures to improve the quality of the artificially upscaled images. The neural networks trained in this thesis can be utilized in the process of acquiring higher quality data from low resolution electron microscope images.
|
|
|
Robust Audio Dereverberation and Denoising
Košina, Simon ; Skácel, Miroslav (referee) ; Szőke, Igor (advisor)
The goal of this thesis was to create a speech enhancement and dereverberation model for audio recordings coming from aircraft VHF communication. First, the thesis covers some theoretical grounds of machine learning and types of neural networks commonly used in such scenarios. Following is a description of the used framework, datasets and the implementation itself. Last chapters are focused on the performed experiments and their evaluation. At the end we talk about the future work that can be done in order to further improve the achieved results.
|
|
|
GaN deposition on a tungsten substrate
Pikna, Štěpán ; Piastek, Jakub (referee) ; Čalkovský, Vojtěch (advisor)
This bachelor thesis is focused on deposition of GaN nanocrystals on the etched tungsten tips. Motivation was to prepare these GaN structures on the Schottky cathode made by company ThermoFisher Scientific and measure its field emission. In the theoretical part of the thesis GaN and tungsten field emission properties are introduced. The experimental part begins with tungsten tip etching optimalization, where the right values for best tips are temperature 20 °C, depth of the tip 2,5 mm and solution NaOH used. Further the gallium structures were prepared on these tips using molecular beam epitaxy (MBE). The right temperature to prepare GaN nanocrystals was determined as 200 °C. The deposition of gallium was set to 2 hours and following nitridation was 3 hours. Finally, the field emission from GaN prepared on copper foil with graphene was measured and compared with other experiments.
|
|
|
Algorithms for improving the detection of selected cardiac arrhythmias
Šandová, Hana ; Ředina, Richard (referee) ; Novotná, Petra (advisor)
The work deals with the generation of ECG arrhythmias that are underrepresented in databases. The theoretical part of the thesis is devoted to a literature search of academic publications that deal with the classification of arrhythmia by using deep learning and data augmentation metod for ECG. The practical part of the thesis deals with noise generator, because adding noise to signals could make the dataset richer. Functions for augmentation of atrial flutter and 3rd and 2nd atrioventricular block were created. It has been tried generation of 2nd atrioventricular block using generative adversarial networks (GAN). Deep learning-based ECG classifiers were used for evaluating the efficiency of the proposed technique in generating synthetic ECG data.
|
|
|
Learning the Face Behind a Voice
Kyjonka, Mojmír ; Matějka, Pavel (referee) ; Plchot, Oldřich (advisor)
This thesis deals with face reconstruction based on voice. The state of the art of this problem is investigated and model for such problem is trained. Model used in this thesis is based on the work "Reconstructing faces from voices" which architecture is based on Generative Adversarial Network (GAN). In this work, we used VGGFace and VoxCeleb datasets, and additionally, we created a small audiovisual dataset of Czech speakers. This work was implemented using the Python scripting language and PyTorch library.
|
|
|
Universal Power Sequencer For Rf Power Amplifiers
Waldecker, Miroslav
More often used Galium Nitride (GaN) based Radio-Frequency high power transistors in the various RF PA configurations e.g. Doherty is by their nature easily destroyed, great care must be taken, when powering-up and shutting down this circuits. That means, proper power biasing and sequencing is necessary. The Doherty type RF PA with RF drivers four different gate, drain voltages and time when the individual voltages are turned on or off must be controled. Universal power sequencer and biasing device, which meets this requirements is described in this article.
|
|
|
Simulation of Skin Diseases Effect Using GAN
Bak, Adam ; Goldmann, Tomáš (referee) ; Kanich, Ondřej (advisor)
Cieľom tejto diplomovej práce je vygenerovanie datasetu syntetických snímkov odtlačkov prstov, ktoré vykazujú známky kožných ochorení. Práca sa zaoberá poškodením spôsobeným kožnými ochoreniami v odtlačkoch prstov a generovaním syntetických odtlačkov prstov. Odtlačky prstov s prejavom kožných ochorení boli generované s využitím modelu založeného na Wasserstein GAN s penalizáciou gradientu. Na trénovanie GAN modelu bola použitá unikátna databáza odtlačkov prstov s prejavom kožných ochorení vytvorená na FIT VUT. Daný model bol trénovaný na troch typoch kožných ochorení: atopický ekzém, psoriáza a dyshidrotický ekzém. Sieť generátoru z natrénovaného WGAN-GP modelu bola použitá na vygenerovanie datasetov syntetických odtlačkov prstov. Tieto syntetické odtlačky boli porovnané s reálnymi odtlačkami s využitím NFIQ a FiQiVi nástrojov na určenie kvality spoločne s porovnaním rozložení lokácií a orientácii markantov v snímkoch odtlačkov prstov.
|
guest ::
