|
|
Methods for Realtime Voice Deepfakes Creation
Alakaev, Kambulat ; Pleško, Filip (referee) ; Malinka, Kamil (advisor)
Tato práce zkoumá možnosti generování hlasových deepfake v reálném čase pomocí nástrojů s otevřeným zdrojovým kódem. Experimenty bylo zjištěno, že rychlost generování hlasových deepfakes je ovlivněna výpočetním výkonem zařízení, na kterých jsou nástroje pro tvorbu řeči spuštěny. Byl identifikován model hlubokého učení, který je schopen generovat řeč téměř v reálném čase. Omezení nástroje obsahujícího tento model však bránila kontinuálnímu zadávání vstupních dat pro generování v reálném čase. K řešení tohoto problému byl vyvinut program, který tato omezení překonává. Kvalita generovaných deepfakes byla hodnocena jak pomocí modelů pro detekci hlasových deepfake, tak pomocí online průzkumů na lidech. Výsledky ukázaly, že zatímco model dokázal oklamat detekční modely, nebyl úspěšný při oklamání lidí. Tento výzkum upozorňuje na dostupnost nástrojů pro syntézu hlasu s otevřeným zdrojovým kódem a na možnost jejich zneužití jednotlivci k podvodným účelům.
|
|
|
System for Recognizing Disinformation in Web Environment
Večerka, Lukáš ; Žádník, Martin (referee) ; Strnadel, Josef (advisor)
This work deals with the design, implementation, and verification of a system for automatic recognition of disinformation on the web. It addresses the issue of disinformation spread in the online environment and its impact on society. It focuses on training several Czech transformer language models for disinformation recognition and further automatic extraction of content from Czech online newspapers and their analysis using text classification and natural language processing through deep learning methods. The results of these analyses are then presented in a web user interface with the aim of providing a platform for verifying articles, authors, and sources. The interface could be used for data annotation by experts for continuous improvement of language models.
|
|
|
Machine Learning of Representations in Genetic Programming
Pomykal, Šimon ; Piňos, Michal (referee) ; Sekanina, Lukáš (advisor)
The aim of this thesis is to become acquainted with machine learning methods that are used for the automatic design of representations. Specifically, the work focuses on deep learning in the field of genetic programming (GP). Image processing is chosen as a case study, particularly noise reduction methods. By combining the acquired knowledge, a new representation is proposed, intended to replace the syntactic tree in the GP algorithm. This method is obtained using a transformer-type neural network. In conclusion, a modified version of GP that works with the new representation is created. This variant is compared with the original GP using the traditional representation in several experiments.
|
|
|
Image Inpainting using Deep Learning
Zobaník, Radek ; Kubík, Tibor (referee) ; Šilling, Petr (advisor)
In this thesis, an application was developed for testing and comparing methods for completing missing parts of an image using deep learning, and two methods were trained, pconv with convolutional architecture, and AOT-GAN with GAN architecture. The thesis describes the design of the finished application, its functionality, and important implementation details. A dataset was selected on which the chosen models were optimally trained. Experiments were made on the AOT-GAN model to investigate the impact of the number of AOT blocks in generator on the resulting completed image. All experiments were qualitatively and quantitatively compared. The results showed respectable outcomes when working with natural scenery.
|
|
|
Atrial fibrillation localization for burden assessment
Martinásková, Klára ; Ředina, Richard (referee) ; Filipenská, Marina (advisor)
The diploma thesis deals with the problem of detection of atrial fibrillation from ECG recordings and localization of given fibrillation segments in signals with paroxysmal fibrillation. A research is done on atrial fibrillation, the origin of this pathology and methods of fibrillation detection from ECG recordings using deep learning. Subsequently, a convolutional neural network model with residual blocks is implemented in Python to classify short (3 s) segments of the ECG signal. Subsequently, the classification results are processed and the segments with paroxysmal fibrillation are localized in the signals with fibrillation. With the classification and localization, the burden assessment of fibrillation is further evaluated. The implemented classifier on the test set achieves an F1 score of 96,15 %. When the sections with fibrillation are localized by the algorithm, MAE of 0,95 s for detecting the beginnings and 1,29 s for detecting the ends with respect to the reference positions is achieved. The estimated patient's burden assessment is compared with the actual values and achieves MAE of 3 %
|
|
|
Signature verification using neural network-based algorithms
Čírtek, Petr ; Kiac, Martin (referee) ; Myška, Vojtěch (advisor)
Signature is one of the most used biometrics in banking and contracting therefore is important to verificate signature authenticity. Verification can be done with the help of a forensic specialist or, thanks to the rise of advanced technology, with the help of a computing technology. The purpose of this thesis is to develop methods for signature verification using neural networks for Czech type of signature and to find out if adding manual extracted features to convolutional analysis could improve these methods. Neural networks seek to replicate the functioning of human brain, consisting of input neurons, several hidden layers and output neurons. Neural networks are one of the most popular artificial intelligence technologies for image analysis and classification. The proposed methods in this thesis work on the principles of convolutional networks. The first proposed method consist of three convolutional layers which extract important features from image of signature and pass them to fully connected classifier layer. This determines whether the signature is genuine or forgery. Also for this method there were created two functions which can interpret it's decision-making. The second method, siamese neural network, unlike the first, does not work with signatures independently, but uses a reference signature image to determine authenticity. The basis of this method is to extract features with convolutional analysis from both the reference signature and the signature to be authenticated. These features are then concatenated and passed to the clasificator. A Czech dataset was created to train models that would verify the Czech type of signatures. From the experiments, it was found that the addition of manualy extracted features has the potential to improve the prediction accuracy of methods based on convolutional image analysis. 3 models were trained, which can verify the Czech type of signatures with an accuracy higher than 80 \%, namely: the model of the convolutional neural network method with discrete wavelet transformation feature, which was trained on the Czech dataset, the model of the same method trained on the CEDAR dataset with number of strokes as added feature and a siamese convolutional neural network method model trained on the Czech dataset of signatures with the tri-surface feature.
|
|
|
Evolutionary Design of Non-Linear Functions for Convolutional Neural Networks
Hladiš, Martin ; Mrázek, Vojtěch (referee) ; Sekanina, Lukáš (advisor)
The aim of this thesis is to design and implement a program for automated design of nonlinear activation functions for convolutional neural networks (CNN) using evolutionary algorithms. The use of automated design provides an independent view to systematically explore a wide range of activation functions and identify the best ones. The method for automatic design chosen in this thesis is a form of evolutionary algorithms referred to as Cartesian genetic programming, which uses a graph representation to encode the solution. This technique allows for the definition of a set of mathematical primitives that define the search space, and thus simply parameterize the design. The implemented approach has been tested on several different architectures and datasets (LeNet-5 \& MNIST, ResNet-10 \& FashionMNIST, WRN-40-4 \& CIFAR-10). Experiments have shown that the approach can find activation functions that statistically improve the accuracy of the architecture over the commonly used ReLU function.
|
|
|
Retinal Images Generation with a Limited Amount of Training Data
Senichak, Yahor ; Semerád, Lukáš (referee) ; Kavetskyi, Andrii (advisor)
The purpose of this study is to explore the progress and application of computer vision and generative adversarial networks (GANs3.1) in the diagnosis and study of fundus diseases. Particular attention is paid to the latest advances in the field of medical data synthesis and the development of our own algorithm. Recent advances in the deep learning architecture U-GAT-IT [22], which includes two pairs of deep neural networks (two generators and two discriminators), have been implemented. This implementation was trained for approximately 300,000 iterations, during which positive results were obtained. The dynamics of the training process were recorded and tests were performed to demonstrate the ability to generate high-quality synthetic images of the ocular background independent of the input data
|
|
|
Deep learning-based noise reduction in X-ray images
Říhová, Barbora ; Jakubíček, Roman (referee) ; Zemek, Marek (advisor)
Technologie zobrazování pomocí rentgenových paprsků je základem zkoumání vnitřní struktury velké škály objektů a výsledky mohou být právě kvůli šumu kompromitovány. Tato práce se zabývá odstraňováním šumu v rentgenových projekcích pomocí hlubokého učení, které má schopnost adaptovat se na konkrétní problém. Práce obsahuje teoretickou rešerši zaměřenou na oblasti produkce a detekce rentgenových paprsků, šumu v rentgenových snímcích a neuronových sítí. Speciální kapitola je věnována popisu vybraného řešení, které je provedeno pomocí tvorby datasetu složeného z části z modelovaných rentgenových projekcí s následně implementovaným šumem odpovídající modelu v reálných snímcích a částečně ze sérií rentgenových projekcí získaných ze zařízení Rigaku nano3DX. K implementaci byla vybrána architektura konvoluční neuronové sítě RIDNet, vzhledem k tomu, že poskytuje v oblasti redukce šumu dobré výsledky. Byly natrénovány tři modely s použitím různých částí datasetu. Nejlepší výkon byl pozorován u modelů, u kterých byla při trénování použita reálná data. Jejich účinnost je srovnatelná s tradičními metodami jako BM3D.
|
|
|
Comparison of Methods for Image Inpainting based on Deep Learning
Rajsigl, Tomáš ; Herout, Adam (referee) ; Španěl, Michal (advisor)
This bachelor thesis aims to compare deep learning methods and approaches for image inpainting using quantitative metrics like PSNR, SSIM, and LPIPS. Moreover, a user study has also been carried out for further subjective assessment. For the purposes of this comparison, four GAN-based neural networks were used. The first network, AOT-GAN, represents a benchmark against which the proposed architecture and its modifications were compared. In the experiments, a variant of the proposed method achieved a 29% improvement against AOT-GAN in images with small missing regions. This claim is also supported by the results of the user study where this method was ranked as the best. As a result of this thesis, a small dataset specifically for the evaluation of image inpainting in the context of object removal was created. Real-world applications of these methods are demonstrated through a web application.
|
guest ::
