|
|
Modelling Music Waveforms Using Wavenet
Slanináková, Terézia ; Landini, Federico Nicolás (referee) ; Beneš, Karel (advisor)
This thesis focuses on exploring the possibilities of modelling music and speech with WaveNet, a deep neural network for generating raw audio waveforms. Using existing implementations, WaveNet was trained on multiple datasets and produced several audio files. Multiple experiments were carried out with various hyperparameter setups of WaveNet to find the optimal settings for the best results. Furthermore, multiple generation schemes were used, each having varying impact on the quality of generated audio. This quality was evaluated using human assessment via a questionnaire, where the musical samples were rated with a score 2-3.1818 on a 5 point scale, which is comparable to the rating of referential audio from the original WaveNet paper (3.1818).
|
|
|
Deep Neural Network Optimization
Bažík, Martin ; Wiglasz, Michal (referee) ; Sekanina, Lukáš (advisor)
The goal of this thesis was to design, implement and analyze various optimizations of deep neural networks, in order to improve the observed parameters. The optimizations are based on modification of the data representation used by neural network operations and searching for the best combination of its hyper-parameters. The convolutional neural networks used for these optimizations were built on LeNet-5 architecture and trained on MNIST, CIFAR-10, and SVHN datasets. The neural networks and their optimizations were implemented within Tiny-dnn library using C++ programming language.
|
|
|
Generation of Authentic Latent Fingerprints Background
Gajda, Adam ; Goldmann, Tomáš (referee) ; Kanich, Ondřej (advisor)
This bachelor's thesis deals with the generation of authentic latent fingerprint backgrounds, through the use of deep learning, more specifically with the help of conditional generative adversarial network and other more conventional methods. This work summarizes the basic theoretical information about biometrics including synthetic fingerprints and a introduction into artificial intelligence. The main model proposed in this thesis has not come into fruition due to lack of unique training data. Other possible reasons were discussed. Thus an alternative way of generating latent fingerprint backgrounds was developed and after visual evaluation of the final results and real data the conclusion was positive.
|
|
|
Neural Network Based Dereverberation
Karlík, Pavol ; Černocký, Jan (referee) ; Žmolíková, Kateřina (advisor)
In the past years, the usage of neural networks in speech processing has increased significantly. This thesis focuses on implementing and evaluating a speech dereverberation framework that utilizes a deep neural network (DNN) to estimate the power spectral density of the signal. The proposed framework is based on the state-of-the-art speech enhancement algorithm called Weighted prediction error (WPE), which is known to effectively reduce reverberation from the speech signal. This thesis summarizes the theory of dereverberation, neural networks and the Weighted prediction error algorithm. Different DNN architectures are experimented with and trained using different datasets with varying properties. The results have shown that our framework is able to outperform the conventional WPE, especially in situations where duration of processed signal is short.
|
|
|
Deep Learning for Image Stitching
Držíková, Diana Maxima ; Vaško, Marek (referee) ; Španěl, Michal (advisor)
Zošívanie obrázkov nie je taký neznámy pojem ako sa na prvý pohľad môže zdať. Určite každý bežný používateľ technológií sa už zozámil s pojmom panoramatický obrázok. V pozadí na zariadení sa prekrývajúce sa obrázky zošívajú a tým vzniká vysoko kvalitný obrázok. Na to aby tento proces fungoval, existujúce algorimy musia spoľahlivo a presne detekovať zaujímavé body, podľa ktorých sa dokáže obrázok správne umiesniť. V tejto práci budú predstavené tradičné metódy na zošívanie obrázkov a taktiež aj metódy s pomocou hlbokých neurónových sietí. Hlavné dva modely, ktoré budú opísane a použíté sú implementácie SuperPoint a SuperGlue. Implementácia bude adaptovaná na párovací systém pre viac ako dva obrázky. Ostatné experimenty, ktoré boli vyskúšané a dopomohli k pochopeniu tejto problematiky budú opísane a vyhodnotené.
|
|
|
Video Enhancement Using Convolutional Networks
Skácel, David ; Špaňhel, Jakub (referee) ; Hradiš, Michal (advisor)
Convolutional neural networks (CNN) represent a state-of-the-art approach to non-trivial image processing tasks, including compression artifacts reduction and image super-resolution. As some research groups nowadays show, these networks can also be leveraged to perform such tasks on real-world video data, resulting in video spatial super-resolution and more. The main goal of this work is to determine whether these nets can be adjusted to perform temporal super-resolution of real-world video data. I utilize the aforementioned neural net architectures in this paper to do so. As I show, given that the input videos are of reasonable quality, these nets are capable of double-image interpolation up to a certain level, where the output image is usable for temporal upsampling. Although the presented results are promising, I encourage more research to be done on this topic.
|
|
|
Deep Learning for Facial Recognition in Video
Jeřábek, Vladimír ; Sochor, Jakub (referee) ; Hradiš, Michal (advisor)
This work deals with face recognition in video using neural networks. In the beginning, there is described the process of selection and verification of convolution neural network to generate feature vectors from images of different identities. In the next part, this work deals with the aggregation of feature vectors from video frames. Aggregation takes place through aggregation neural networks. At the end of this work, the results obtained by the aggregation methods are discussed.
|
|
|
Deep Learning for Facial Recognition in Video
Stratil, Jan ; Sochor, Jakub (referee) ; Hradiš, Michal (advisor)
This bachelor's thesis deals with facial recognition in video using deep neural networks. This task is split into 2 parts. The first part deals with training network that produces compact feature vector which represents the face identity from a video frame. The second part deals with training aggregation network that aggregates those feature vectors into one. This aggregation is fast and it has shown that its results are better than naive pooling methods. Results are tested on the LFW dataset, where it achieves 92.8% accuracy and on the YTF dataset, where the accuracy is 84.06%.
|
|
|
Deep Learning for Image Classification
Ziková, Jana ; Veľas, Martin (referee) ; Hradiš, Michal (advisor)
This bachelor thesis deals with electronic commerce website products classification using product's photographs. For this purpose we use already implemented models of deep convolutional neural networks. Tho goal of this theses is to design experiments that will lead to the best possible results in product images classification.
|
|
|
Multi-Task Neural Networks for Speech Recognition
Egorova, Ekaterina ; Veselý, Karel (referee) ; Karafiát, Martin (advisor)
První část této diplomové práci se zabývá teoretickým rozborem principů neuronových sítí, včetně možnosti jejich použití v oblasti rozpoznávání řeči. Práce pokračuje popisem viceúkolových neuronových sítí a souvisejících experimentů. Praktická část práce obsahovala změny software pro trénování neuronových sítí, které umožnily viceúkolové trénování. Je rovněž popsáno připravené prostředí, včetně několika dedikovaných skriptů. Experimenty představené v této diplomové práci ověřují použití artikulačních characteristik řeči pro viceúkolové trénování. Experimenty byly provedeny na dvou řečových databázích lišících se kvalitou a velikostí a representujících různé jazyky - angličtinu a vietnamštinu. Artikulační charakteristiky byly také kombinovány s jinými sekundárními úkoly, například kontextem, s záměrem ověřit jejich komplementaritu. Porovnaní je provedeno s neuronovými sítěmi různých velikostí tak, aby byl popsán vztah mezi velikostí neuronových sítí a efektivitou viceúkolového trénování. Závěrem provedených experimentů je, že viceúkolové trénování s použitím artikulačnich charakteristik jako sekundárních úkolů vede k lepšímu trénování neuronových sítí a výsledkem tohoto trénování může být přesnější rozpoznávání fonémů. V závěru práce jsou viceúkolové neuronové sítě testovány v systému rozpoznávání řeči jako extraktor příznaků.
|
guest ::
