|
|
Aplikace pro záznam fotek a videoklipů ovládaná hlasem
Jurčík, Adam Dalibor ; Španěl, Michal (oponent) ; Herout, Adam (vedoucí práce)
Cílem této práce je vytvořit aplikaci, která řeší problémy spojené s nahráváním videa a fotografií téže osoby při sportovních aktivitách. Tyto problémy jsou například, potřeba mít u sebe dálkové ovládání nebo neustálé přerušování činnosti z důvodu ručního spuštění. Lze situaci vyřešit pomocí hlasového ovládání a klíčových slov, po kterých se spustí pořízení fotografie. V práci je podrobně rozebráno, jakým způsobem lze implementovat hlasové ovládání na operačním systému Android v programovacím jazyce Java dohromady s pořízením snímku či videa. Dále obsahuje vysvětlení ukládání těchto souborů a to, jak ovládat hardware zařízení jako například fotoaparát. Práce obsahuje podrobnější popis dvou přístupů k rozpoznání hlasu a popisuje jejich funkčnost v různých způsobech využití.
|
|
|
Automatic Tire Inspection Using Surface Scans
Toth Vaňo, Pavol ; Materna, Zdeněk (oponent) ; Španěl, Michal (vedoucí práce)
This thesis deals with automatic detection of defects on tire treads using their depth scans. The approach proposed in the thesis doesn’t require a faultless reference tire for the inspected tire. The first step is the detection of anomalies, which is done using a modification of the PatchCore method proposed in the thesis, taking advantage of the repetition of patterns on the tire tread. Subsequently, anomalies corresponding to special elements on the tire are detected using the deep neural networks Faster R-CNN and Deep Hough transform, and they are filtered out. Applying the proposed approach on the prepared dataset, the value 0.584 of Average Precision metric for detection was obtained. The biggest weakness of the proposed method is its limited ability to detect defects with a very small depth.
|
|
|
Aplikace pro odstranění objektů z obrázků s využitím metod hlubokého učení
Kotoun, Josef ; Hradiš, Michal (oponent) ; Španěl, Michal (vedoucí práce)
Práce se zabývá tvorbou webové aplikace, která uživateli umožňuje jednoduchým způsobem vybrat objekt v obrázku a následně ho odstranit vizuálně věrohodným způsobem. Aplikace je implementována s použitím frameworku SvelteKit. K výběru a odstranění objektů jsou využity neuronové sítě Mobile Segment Anything a Mobile Inpainting GAN. Neuronové sítě jsou spouštěny na klientské straně webové aplikace s využitím knihovny ONNX Runtime Web, pro efektivní využití zdrojů klientova zařízení jsou použity technologie WebGPU a WebAssembly. Výsledná aplikace umožňuje díky využitým neuronovým sítím vybrat a odstranit objekt pouze v několika málo kliknutích. Dle zpětné vazby uživatelů se s aplikací pracuje jednoduše a upravená část výsledné fotografie je ve většině případů jen těžko postřehnutelná.
|
|
|
Deep Neural Networks for Landmark Detection in Images
Stolárik, Samuel ; Španěl, Michal (oponent) ; Kubík, Tibor (vedoucí práce)
Landmark detection in medical images is essential for the diagnosis of various medical conditions. However, it remains laborious, time-consuming, and therefore costly. Deep neural networks have proved themselves to be an applicable solution for the creation of fully automatic landmark detection systems. Automatic landmark detection in the medical domain faces the obstacle of insufficient training data; because of this, using the most efficient representation of the data is important. In this work, three different approaches to the representation of data during the training of deep neural networks were tested. The first two use the U-Net model for heatmap regression and differ in the way how the heatmaps are created. One uses 2D Gaussian heatmaps, which, however, ignore the underlying anatomical information in the input image, and the second one addresses this problem by using novel contour-hugging heatmaps. The third approach uses a graph neural network to process the hierarchical graph representation of the input image. Experiments demonstrated that the combination of the U-Net and the novel contour-hugging heatmaps could detect landmarks in cephalometric images with a mean error of 1.8 mm. This represents a 3.34% improvement over the 2D Gaussian heatmaps. Additionally, the method predicted 70.21% of landmarks with an error of less than 2 mm.
|
|
|
Porovnání metod pro doplnění chybějící části obrazu založených na hlubokém učení
Rajsigl, Tomáš ; Herout, Adam (oponent) ; Španěl, Michal (vedoucí práce)
Cílem této bakalářské práce je porovnat metody hlubokého učení pro doplnění chybějící části obrazu pomocí kvantitativních metrik jako jsou PSNR, SSIM a LPIPS. Pro dodatečné subjektivní ohodnocení byla taktéž provedena uživatelská studie. K porovnání byly použity celkem čtyři neuronové sítě založené na architektuře GAN. Navrhovaná architektura neuronové sítě a její modifikované verze byly porovnávány oproti síti AOT-GAN. Experimenty ukázaly, že v obrazech s malou chybějící částí dosáhla varianta navržené metody 29% zlepšení oproti již zmiňované metodě AOT-GAN. Toto tvrzení podporují i výsledky uživatelské studie, kde byla tato metoda vyhodnocena jako nejlepší. V rámci této práce vznikla malá datová sada určená pro vyhodnocení metod retušování obrazu při úloze odstraňování objektů. Reálné využití těchto metod je demonstrováno prostřednictvím webové aplikace.
|
|
|
Stereo Reconstruction with Deep Neural Networks
Letanec, Richard ; Herout, Adam (oponent) ; Španěl, Michal (vedoucí práce)
The aim of this thesis is to design and train a neural network model capable of estimating a disparity map from a pair of images. It will then be possible to create a depth map and point cloud from the estimated disparity map. Such a process is called stereo reconstruction. Solving this task consists of two steps -- choosing a suitable dataset and choosing a suitable neural network architecture. In my work, I compared two neural network architectures that I trained on the DrivingStereo dataset, consisting of paired images photographed from the roof of a car, and retrained and evaluated on the KITTI 2015 dataset, consisting of images of the same type. As the first neural network architecture, I chose ES-Net, which uses an approach based on a sequence of residual blocks and convolutional layers. As the second architecture, I chose CREStereo, which uses an iterative approach based on recurrent layers to predict the disparity map. In all benchmark tests, the CREStereo architecture achieves better accuracy.
|
|
|
Video Denoising Using Deep Learning
Naumenko, Maksim ; Hradiš, Michal (oponent) ; Španěl, Michal (vedoucí práce)
In the era of digital multimedia, video content quality significantly impacts user experiences and system performance, particularly in domains such as entertainment, and video and image processing. This thesis addresses the persistent challenge of video noise, which degrades video quality, through the use of advanced deep learning techniques. Initially, traditional video denoising approaches are reviewed to establish a foundational understanding of denoising concepts. Subsequently, two state-of-the-art models, FastDVDNet and ViDeNN, are studied to familiarize with neural network architectures. The main product of this work is the development of a robust video denoising pipeline that utilizes a UNet architecture inspired by these state-of-the-art models. Throughout the thesis, the proposed UNet Baseline, ResUNet, and ResUNet Temporal models are explained, implemented, and evaluated to demonstrate their effectiveness in video denoising.
|
|
|
Deep Learning for 3D Mesh Registration
Pukanec, Dávid ; Beran, Vítězslav (oponent) ; Španěl, Michal (vedoucí práce)
The problem of mesh alignment is often solved through point cloud registration. Numer- ous deep learning-based registration methods are published every year achieving state-of- the-art results. Based on their core concepts, the methods can loosely be divided into correspondence-based and correspondence-free. Even though comparisons of individual methods exist, the cross-evaluations of both categories are lacking. In this work, a deeper evaluation of Lepard and FINet models is presented. For this purpose, the ModelNet40 and Teeth3DS datasets are used. The experiments show that FINet is able to align unseen shapes, obscured by partiality and noise with a translation error of 4.16% of model size and a rotation error of 3.640 degrees. While Lepard manages this with a translation error of 6.73% of model size and a rotation error of 7.265 degrees.
|
|
|
Generative Models for 3D Shape Completion
Zdravecký, Peter ; Španěl, Michal (oponent) ; Kubík, Tibor (vedoucí práce)
In many real-world scenarios, scanned 3D models contain missing parts due to occlusion, scanning errors, or the incomplete nature of the data itself. The goal of this work is to create an automated process for 3D shape completion using a supervised deep learning-based method. The proposed solution is based on the prior work of DiffComplete, which uses a diffusion-based model operating over distance field representation and handles the task as a generative problem. The results showed a high capability of this model with an 81.6 IoU metric on the custom-prepared test set of furniture objects. The model also demonstrates strong generalization capabilities on shapes that are out of the training distribution (average 70.9 IoU metric). Apart from more detailed data-centric experiments, this work further extends current state-of-the-art in two ways. Firstly, it addresses the most crucial shortcoming, expensive computation, by processing the input in a low-resolution domain. Secondly, it utilizes user input (Region of Interest), which gives the user more control over generation in ambiguous scenarios.
|
|
|
Deep Learning for Image Stitching
Šilling, Petr ; Beran, Vítězslav (oponent) ; Španěl, Michal (vedoucí práce)
Image stitching is an essential technique for reconstructing volumes of biological samples from overlapping tiles of electron microscopy (EM) images. Current volume EM stitching methods generally rely on handcrafted features, such as those produced by SIFT. However, recent developments indicate that convolutional neural networks (CNNs) can improve stitching accuracy by learning discriminative features directly from training images. Taking into account the potential of CNNs, this thesis proposes DEMIS, a novel EM image stitching tool based on LoFTR, an attention-based feature matching network. The thesis also proposes a novel dataset generated by splitting high-resolution EM images into grids of overlapping image tiles. The dataset is used to fine-tune LoFTR and to evaluate the DEMIS tool. Experiments on the synthetic dataset reveal higher feature matching accuracy compared to SIFT. Moreover, experiments on challenging images with small overlap regions and high resolution demonstrate significantly higher stitching robustness than SIFT. Overall, the results suggest that deep learning methods could be beneficial for EM imaging, for example, by allowing the use of smaller tile overlaps.
|
host ::
RSS.