|
|
Neural Networks for Video Quality Enhancement
Sirovatka, Matej ; Juránek, Roman (referee) ; Hradiš, Michal (advisor)
Cieľom tejto práce je vytvoriť novú metódu super rozlíšenia na zlepšenie kvality videa. Táto metóda je založená na myšlienke použitia deformovateľných konvolučných vrstiev a optického toku na zarovnanie príznakov z viacerých po sebe následujúcích snímkov videa. Táto metóda je následne použitá v neuronovej sieti založenej na U-Net architektúre na predikciu snímkov vo vysokom rozlíšení. Vyhodnotenie je prevedené na datasete obsahujúcom snímky z reálneho života a porovnané s inými metódami. Testované sú rôzne konfigurácie navrhnutej metódy a výsledky sú analyzované. Výsledky experimentov ukazujú sľubné výsledky, pričom model prekonáva bilineárnu interpoláciu a metódy založené na jednom snímku. Testované sú rôzne architektúry modulu zarovnávania príznakov spolu s celou architektúrou U-Net, pričom sa ukazuje, že použitie Vgg19 ako enkóderu dáva najlepšie výsledky.
|
|
|
Interaktivní nástroj pro bike fitting využívající počítačové vidění
Kocman, Matej ; Hradiš, Michal (referee) ; Beran, Vítězslav (advisor)
Bike fitting is the adjustment of a cyclist’s pose on the bike often with the help of video analysis with the aim of increasing comfort, preventing injury and performance optimisation. The goal of this thesis was to create a prototype application for bike fitting, which would follow up on existing applications of its kind and would offer a working solution for chosen problems. The application offers functionality for increased keypoint detection accuracy, lowering the number of position iterations needed for optimal bike setup and optimal pose setup for a bike with limited setup options. The application was tested on users and contributed towards reaching an ideal position after a few iterations already. The application is a useful tool for bike fitting at home and the thesis propositions ways of improving it further.
|
|
|
Automatická vizuální podpora pro Q-řazení
Kán, Dávid ; Hradiš, Michal (referee) ; Vaško, Marek (advisor)
This bachelor thesis deals with the integration of Q-sorting and computer vision methods for object detection. The goal of the work is to create a program that, with the help of~visual support, will facilitate the process and at the same time prevent errors in Q-sorting. Furthermore, the work deals with the creation of~a suitable data set for training the model and for experiments, which takes into account the way the cards are laid out and the~environment. The implemented program takes the form of a console application and is written using the Python programming language. The program uses YOLOv8 to detect objects and uses Pero OCR to retrieve text from cards. Using the created test set, experiments were performed on the trained model and the program was tested.
|
|
|
Application for Objects Removal from Images Using Deep Learning Methods
Kotoun, Josef ; Hradiš, Michal (referee) ; Španěl, Michal (advisor)
The thesis deals with the development of a web application that allows users to easily select an object in an image and then remove it in a visually plausible way. The application is implemented using the SvelteKit framework. Mobile Segment Anything and Mobile Inpainting GAN neural networks are utilized for object selection and removal. The neural networks are executed on the client-side of the web application using the ONNX Runtime Web library. To efficiently utilize client device resources, WebGPU and WebAssembly technologies are employed. Thanks to the neural networks used, the resulting application enables users to select and remove objects in just a few clicks. According to user feedback, the application is easy to use, and the edited part of the resulting photograph is barely noticeable in most cases.
|
|
|
Webová aplikace pro efektivní anotaci atributů objektů ve videu
Pernický, Michal ; Kohút, Jan (referee) ; Hradiš, Michal (advisor)
The goal of this work is to develop a web application for annotating video object attributes that combines an efficient user interface with an assistant classifier providing predictions. In contrast to currently available tools, the solution focuses directly on objects without assigning them to the original videos. The ability to filter objects according to their attributes and to confirm or reject predicted attribute values in bulk is important. Testing on users has been found to reduce the time spent working by up to half. This shows that further work with this annotation principle is worthwhile.
|
|
|
Video Denoising Using Deep Learning
Naumenko, Maksim ; Hradiš, Michal (referee) ; Španěl, Michal (advisor)
V éře digitálních multimédií kvalita videoobsahu významně ovlivňuje uživatelský zážitek a výkon systému, zejména v oblastech, jako je zábava a zpracování videa a obrazu. Tato práce se zabývá přetrvávajícím problémem šumu ve videu, který zhoršuje jeho kvalitu, a to pomocí pokročilých technik hlubokého učení. Nejprve jsou přezkoumány tradiční přístupy k odstraňování šumu ve videu, aby bylo možné nastínit základní koncepty denoisingu. Následně jsou studovány dva referenční modely, FastDVDNet a ViDeNN, za účelem seznámení se s architekturami neuronových sítí. Hlavním výsledkem této práce je vývoj robustního systému pro odstraňování šumu ve videu, který je založen na architektuře UNet inspirované těmito referenčními modely. V průběhu práce jsou vysvětleny, implementovány a vyhodnoceny navrhované modely UNet Baseline, ResUNet a ResUNet Temporal, aby byla prokázána jejich účinnost v odstraňování šumu ve videu.
|
|
|
Guided Reinforcement Learning for Motor Skills
Karabelly, Jozef ; Herout, Adam (referee) ; Hradiš, Michal (advisor)
Cieľom tejto práce je prezentovať prehľad aktuálneho výskumu v oblasti posilovaného učenia pohybu s predlohou a identifikovať potenciálne smery výskumu. Okrem toho práca predstavuje vylepšenú metódu učenia fyzikálne simulovaných animácií postáv založenú na aktuálnych metódach. Predtrénovaný model ukazuje potenciál lepších výsledkov na rôznych nových úlohách. Vlastný dataset bol nazbieraný pre účely pretrénovania modelu predstaveného v tejto práci. Na základe výsledkov z vykonaných experimentov sú odprezentované možné budúce vylepšenia a smery výskumu.
|
|
|
Page Layout Analysis with Graph Neural Networks
Otčenáš, Matej ; Kišš, Martin (referee) ; Hradiš, Michal (advisor)
The aim of this work is to experimentally test the power of graph neural networks in the comprehensive analysis of document layout. In terms of document types, the focus is primarily on newspaper articles and historical writings, such as handwritten books or medieval manuscripts. These are characterized by the complexity of their layout, lacking a fixed structure or having highly segmented text. The work deals with the creation of suitable datasets for training and testing an approach for globally ordering the sequence of reading lines on a page and assigning each line to one of the defined classes. The research also involves creating an appropriate representation of a graph that captures relationships between individual components on the page and selecting a suitable graph neural network with the appropriate parameters. Finally, the different approaches are evaluated and compared on multiple metrics suitable for the given problem, and the findings are summarized with a discussion on possible enhancements and limitations.
|
|
|
Detekce objektu s využitím hloubkových dat
Valko, Marek ; Hradiš, Michal (referee) ; Musil, Petr (advisor)
This bachelor thesis addresses the detection of objects in images using depth data. The goal was to select appropriate deep learning methods and experimentally verify them on relevant datasets. The thesis begins with an overview of basic techniques for detecting objects in images and depth data, utilizing selected datasets NYU Depth v2 and Washington RGB-D to test modified YOLOv5 and YOLOv8 models, adapted for effective processing of RGB-D data. The experiments explored various representations of depth information and analyzed how the integration of depth data enhances the performance of these models. The results demonstrated significant improvements in mAP metrics compared to traditional models that use only RGB data. The integration of depth data thus allowed for more accurate and reliable object detection results.
|
|
|
Playing Games Using Neural Networks
Buchal, Petr ; Kolář, Martin (referee) ; Hradiš, Michal (advisor)
The aim of this bachelor thesis is to teach a neural network solving classic control theory problems and playing the turn-based game 2048 and several Atari games. It is about the process of the reinforcement learning. I used the Deep Q-learning reinforcement learning algorithm which uses a neural networks. In order to improve a learning efficiency, I enriched the algorithm with several improvements. The enhancements include the addition of a target network, DDQN, dueling neural network architecture and priority experience replay memory. The experiments with classic control theory problems found out that the learning efficiency is most increased by adding a target network. In the game environments, the Deep Q-learning has achieved several times better results than a random player. The results and their analysis can be used for an insight to reinforcement learning algorithms using neural networks and to improve the used techniques.
|
guest ::
