|
| |
|
|
Reconstruction of Missing Parts of the Face Using Neural Network
Marek, Jan ; Drahanský, Martin (oponent) ; Goldmann, Tomáš (vedoucí práce)
The goal of this thesis is to design a neural network for reconstruction of face images in which a part of the face is obscured by a mask. Concepts used in the development of convolutional neural networks and generative adversarial networks are presented. Specific concepts used in neural networks used for face reconstruction are described. The generative adversarial network presented in this thesis combines the use of gated convolutional layers and dense multiscale fusion blocks to produce realistic reconstructions of masked face images.
|
|
|
Synthetic Fingerprint Generation Using GAN
Dvořák, Jiří ; Drahanský, Martin (oponent) ; Kanich, Ondřej (vedoucí práce)
This thesis is focused on the generation of synthetic fingerprints using a model based on the principle of generative adversarial networks. The work summarizes the basic theoretical information about biometrics with emphasis on fingerprints. It also describes the principle of one of the popular synthetic fingerprint generators called SFinGe. The model based on a deep convolutional generative adversarial network is discussed together with several methods that improved its performance. The results were evaluated by computing the Fréchet Inception Distance between the generated and real fingerprints. The generated dataset of 100 samples was also evaluated by NFIQ 2.0 which proved that the proposed model is able to generate fingerprints with almost the same quality of the training samples.
|
|
|
Speech Enhancement with Cycle-Consistent Neural Networks
Karlík, Pavol ; Černocký, Jan (oponent) ; Žmolíková, Kateřina (vedoucí práce)
Deep neural networks (DNNs) have become a standard approach for solving problems of speech enhancement (SE). The training process of a neural network can be extended by using a second neural network, which learns to insert noise into a clean speech signal. Those two networks can be used in combination with each other to reconstruct clean and noisy speech samples. This thesis focuses on utilizing this technique, called cycle-consistency. Cycle-consistency improves the robustness of a network without modifying the speech-enhancing neural network, as it exposes the SE network to a much larger variety of noisy data. However, this method requires input-target training data pairs, which are not always available. We use generative adversarial networks (GANs) with cycle-consistency constraint to train the network using unpaired data. We perform a large number of experiments using both paired and unpaired training data. Our results have shown that adding cycle-consistency improves the models' performance significantly.
|
|
|
Generation of Synthetic Retinal Images with High Resolution
Aubrecht, Tomáš ; Heidari, Mona (oponent) ; Drahanský, Martin (vedoucí práce)
Special equipment, a fundus camera, is needed to capture the retina, which is the most important part of the human eye. Therefore, the main objective of this work is to design and implement a system that would be able to generate retinal images. The proposed solution uses an image-to-image translation, where the system is provided with a black and white image at the input containing only bloodstream, on the basis of which a color image of the entire retina is generated. The system consists of two neural networks: a generator, which generates retinal images, and a discriminator, which classifies these images as real or synthetic. Training of this system was performed on 141 images from publicly available databases. A new database was created with more than 2,800 images of healthy retinas in a resolution of 1024x1024. This database could be used as a learning tool for ophthalmologists or for the development of various applications working with retinas.
|
|
|
Automatické kolorování videa
Mikeska, Tomáš ; Kolář, Martin (oponent) ; Hradiš, Michal (vedoucí práce)
Tato práce se zabývá plně automatickým kolorováním videa a fotografií pomocí konvolučních neuronových sítí. Shrnuje dosavadní přístupy, architektury a různé chybové funkce. V rámci práce jsou navrženy a natrénovány různé varianty neuronových sítí s různými chybovými funkcemi pro automatické kolorování. Nejlepší kombinace zvládá kolorovat velké množství scén a je demonstrována její schopnost kolorovat dostatečně koherentní video.
|
|
|
Generating Faces with Generative Adversarial Networks
Konečný, Daniel ; Herout, Adam (oponent) ; Kolář, Martin (vedoucí práce)
The goal of this thesis is generating color images of faces from randomly chosen high-dimensional vectors with Generative Adversarial Networks. The next task is to analyze input vectors based on the features of faces generated from those vectors. Three different models of Generative Adversarial Network are implemented, one for generating images of handwritten digits and other two for generating images of faces. Generated images show credible-looking faces, but recognizable from real ones with a human eye. Single dimensions of input vectors are analyzed with Student's t-test. Linear Discriminant Analysis is then used to project input vectors into subspaces where the classes of features are separable. Analysis of generated data proves that the input vector can be specifically chosen to generate an image of a face with requested features with probability up to 80 %. The main result of this thesis is a model of Generative Adversarial Network for generating images of faces. A tool for generating images of faces with chosen features is implemented too.
|
|
|
Příprava trénovacích dat pomocí generativních neuronových sítí
Ševčík, Pavel ; Kolář, Martin (oponent) ; Hradiš, Michal (vedoucí práce)
Cílem této práce byla příprava trénovací datové sady pro detekci dopravních značek pomocí generativních neuronových sítí. V řešení byla použita upravená architektura U-Net a bylo experimentováno s aplikací stylů pomocí vrstev AdaIN podobně jako v modelu StyleGAN. Rozšířením reálné datové sady GTSDB o uměle vytvořené snímky bylo dosaženo úspěšnosti 80,36 %, což představuje zlepšení o 19,27 % oproti úspěšnosti detektoru natrénovanému pouze na reálných datech.
|
|
|
Present State Of GaN Technology In Power Electronics
Šír, Michal
This paper presents a general overview of nowadays Gallium Nitride power transistor technology and shows the existing components with their limits from different manufacturers currently available on the market. Introduction to GaN depletion mode, enhancement mode and cascode transistor structure with their function explanation is included.
|
|
|
Optimizing Bias Point Of High Efficiency Class-B Gan Power Amplifier For The Best Efficiency
Fiser, Ondrej
High performance amplifiers are always a demanding component in the world of wireless communication. The amplifier is the heart that drives each radio system. We have designed and developed a high performance one-stage class-B GaN power amplifier for drone applications in the S-band (at 1,6 GHz) with maximum output power 6 W. This paper compare fixed settings of the bias point option and optimized bias point for the best efficiency within the entire output power range. Applying the proposed method, that is particularly advantageous for low power performance to improve efficiency by more than 15 %.
|
host ::
RSS.