|
|
Coherence gated holographic microscopy
Ďuriš, Miroslav ; Tyc,, Tomáš (oponent) ; Baránek,, Michal (oponent) ; Chmelík, Radim (vedoucí práce)
Biomedical and metasurface researchers repeatedly reach for quantitative phase imaging (QPI) as their primary imaging technique due to its high-throughput, label-free, quantitative nature. Therefore, QPI has quickly established its role in identifying rare events and screening in biomedicine or automated image data analysis using artificial intelligence. These and many other applications share the requirement for extensive high-quality datasets, which is challenging to meet due to obstacles specific to each application. This thesis tackles the principal problems of optical imaging, mainly in biomedical research. The research aimed to study and develop new imaging methods by extending the capabilities of the coherence-controlled holographic microscope. In the thesis, we tackled three principal areas of biomedical imaging: turbid media imaging, super-resolution QPI, and 3D refractive index reconstruction. To achieve such ambitious results, we have utilized the so-called coherence-gating effect, typically exploited for imaging through disordered media by least-scattered (ballistic) light. To tackle turbid media imaging, we counterintuitively use the coherence gate for imaging by the non-ballistic light, enabling us to retrieve information missing in the ballistic image. A combination of images for different coherence gate positions allow us to synthesize an image of quality superior to ballistic light approaches, which we experimentally demonstrate on QPI through thick biological tissue. Two approaches to super-resolution QPI were explored in the thesis. First is the synthetic aperture approach, for which we again exploit the coherence-gating properties of the partially coherent light combined with the oblique illumination provided by the diffraction on a simple hexagonal phase grating placed near the specimen. We synthesize synthetic aperture QPI with significantly increased spatial frequency bandwidth from sequentially acquired images formed by the coherence-gated light scattered into each grating's diffraction order. Second, we developed the coherence gate shaping method allowing real-time super-resolution QPI. We propose an approach based on the fact that our system's point spread function (PSF) is a product of the diffraction-limited spot and the coherence-gating function, which we shape similarly to the superoscillatory hotspot. The product simultaneously produces the PSF with a super-resolution central peak and minimizes sidelobe effects, the common obstacle of superoscillatory imaging. The attenuation of sidelobes and resolution improvement co-occur in the entire field of view. Therefore, we present the first single-shot wide-field super-resolution QPI. For both methods, we achieved a resolution improvement of at least 19\%. Furthermore, we demonstrate the feasibility of the proposed methods by imaging biological specimens with super-resolution. In the thesis, we also address 3D imaging by the coherence-controlled holographic microscope. We developed a method for 3D refractive index distribution reconstruction from a z-stack QPI measurement. The reconstructed refractive index distribution has qualities similar to the outputs of optical diffraction tomography. At the same time, the required number of acquisitions is significantly lower in the case of the proposed method. We demonstrate our approach using simulated as well as experimental data.
|
|
|
Facial image restoration
Bako, Matúš ; Herout, Adam (oponent) ; Hradiš, Michal (vedoucí práce)
In this thesis, I tackle the problem of facial image super-resolution using convolutional neural networks with focus on preserving identity. I propose a method consisting of DPNet architecture and training algorithm based on state-of-the-art super-resolution solutions. The model of DPNet architecture is trained on Flickr-Faces-HQ dataset, where I achieve SSIM value 0.856 while expanding the image to four times the size. Residual channel attention network, which is one of the best and latest architectures, achieves SSIM value 0.858. While training models using adversarial loss, I encountered problems with artifacts. I experiment with various methods trying to remove appearing artefacts, which weren't successful so far. To compare quality assessment with human perception, I acquired image sequences sorted by percieved quality. Results show, that quality of proposed neural network trained using absolute loss approaches state-of-the-art methods.
|
|
|
Enhancement of image quality for security forces
Varga, Adam ; Galáž, Zoltán (oponent) ; Burget, Radim (vedoucí práce)
This bachelor thesis deals with image quality enhancement for security forces. Image quality enhancement in this case means increasing the resolution of image data by using super-resolution techniques using models of deep convolutional neural networks. The thesis in its theoretical part describes the principles of the operation of this technique and in its practical part is presented the work with selected state-of-the-art models in the area of super-resolution.
|
|
|
Image Super-Resolution Using Deep Learning
Mojžiš, Tomáš ; Beran, Vítězslav (oponent) ; Španěl, Michal (vedoucí práce)
The aim of this thesis is to create a deep neural net capable of super-resolution on images acquired by electron microscopes. The thesis consists of two parts - finding appropriate data and creating a dataset for the super-resolution task and designing a neural net architecture capable of solving the super-resolution task. Within the thesis, two datasets comprised of images acquired by electron microscopes were created. The datasets differ in the approach to data augmentation. They allow to train a neural network which fulfills the super-resolution task. To solve this task, two U-Net based and one GAN based architecture were trained. The resolution of images was upscaled by a factor of two and four. The best artificially upscaled images were created by neural network Real-ESRGAN. The values of metrics were not higher than the tested interpolation method, but the images seem more visually pleasing especially when they were upscaled four times. Thanks to this thesis, two datasets were created allowing to train other possible neural network architectures to improve the quality of the artificially upscaled images. The neural networks trained in this thesis can be utilized in the process of acquiring higher quality data from low resolution electron microscope images.
|
|
|
Superresolution
Mezera, Lukáš ; Dvořák, Radim (oponent) ; Orság, Filip (vedoucí práce)
The goal of this thesis is to propose the super-resolution method for the image of the scene when multiple frames of the given scene are available. The theoretical part of this thesis brings the report about current multi-frame super-resolution methods. These methods are compared according to the optimal criteria. The own super-resolution method is proposed in the practical part of this thesis. However this method isn't rotation invariant and for this reason is proposed the improved super-resolution method. There are also suggested improvements of the improved super-resolution method in this thesis.
|
|
|
Determination of Objects Similarity Based on Image Information
Rajnoha, Martin ; Kamencay,, Patrik (oponent) ; Beneš, Radek (oponent) ; Burget, Radim (vedoucí práce)
Monitoring of public areas and their automatic real-time processing became increasingly significant due to the changing security situation in the world. However, the problem is an analysis of low-quality records, where even the state-of-the-art methods fail in some cases. This work investigates an important area of image similarity – biometric identification based on face image. The work deals primarily with the face super-resolution from a sequence of low-resolution images and it compares this approach to the single-frame methods, that are still considered as the most accurate. A new dataset was created for this purpose, which is directly designed for the multi-frame face super-resolution methods from the low-resolution input sequence, and it is of comparable size with the leading world datasets. The results were evaluated by both a survey of human perception and defined objective metrics. A hypothesis that multi-frame methods achieve better results than single-frame methods was proved by a comparison of both methods. Architectures, source code and the dataset were released. That caused a creation of the basis for future research in this field.
|
|
|
Metody zvyšování rozlišení digitálních snímků
Franěk, Pavel ; Fedra, Petr (oponent) ; Mézl, Martin (vedoucí práce)
Cílem této bakalářské práce je se seznámit s metodami, které umožňují zvýšení rozlišení digitálních snímků. Také realizovat jednotlivé interpolační metody i Super-rozlišení pomocí programu Matlab a poukázání na zhodnocené výsledky. Diskutovat o možnostech použití metod Super-rozlišení pro obrazy s lékařských modalit.
|
|
|
Image Super-Resolution Using Deep Learning
Bublavý, Martin ; Juránková, Markéta (oponent) ; Španěl, Michal (vedoucí práce)
The ability to identify and treat a variety of medical diseases is made possible by medical imaging, which is an essential component of contemporary healthcare. Yet, elements like noise and low resolution can have a negative impact on the quality of medical photographs. In this thesis, how to enhance the resolution and quality of medical images was investigated using MedSRGAN, a deep learning model built on generative adversarial networks (GANs). MedSRGAN was implemented and then applied to computed tomography (CT), one of the most utilized medical imaging methods.
|
|
|
Superrozlišení v obraze pro zajištění vylepšeného monitorování zabezpečených prostorů
Rosa, Martin ; Mezina, Anzhelika (oponent) ; Burget, Radim (vedoucí práce)
Cieľom bakalárskej práce bolo porovnať modely super-rozlíšenia s aplikáciou na reštaurovanie fotiek ľudských tvárí. V práci sme spracovali rešerš technológií superrozlíšenia a následne sme natrénovali a porovnali 5 modelov. Zameriavame sa hlavne na oblasť superrozlíšenia, ktorá by mohla byť nápomocná na identifikáciu osôb z bezpečnostných kamier. Použité technológie boli preto vyberané na základe percepčnej kvality a schopnosti identifikácie osoby na výstupnom snímku. Práca ukázala účinnosť porovnaných modelov pomocou objektívnych aj subjektívnych metrík. Výsledky boli porovnané v dotazníku (106 respondentov). Dotazník ukázal účinnosť použitia vlnovej transformácie v superrozlíšení tvarí.
|
|
|
Image Super-Resolution Using Deep Learning
Mojžiš, Tomáš ; Beran, Vítězslav (oponent) ; Španěl, Michal (vedoucí práce)
The aim of this thesis is to create a deep neural net capable of super-resolution on images acquired by electron microscopes. The thesis consists of two parts - finding appropriate data and creating a dataset for the super-resolution task and designing a neural net architecture capable of solving the super-resolution task. Within the thesis, two datasets comprised of images acquired by electron microscopes were created. The datasets differ in the approach to data augmentation. They allow to train a neural network which fulfills the super-resolution task. To solve this task, two U-Net based and one GAN based architecture were trained. The resolution of images was upscaled by a factor of two and four. The best artificially upscaled images were created by neural network Real-ESRGAN. The values of metrics were not higher than the tested interpolation method, but the images seem more visually pleasing especially when they were upscaled four times. Thanks to this thesis, two datasets were created allowing to train other possible neural network architectures to improve the quality of the artificially upscaled images. The neural networks trained in this thesis can be utilized in the process of acquiring higher quality data from low resolution electron microscope images.
|
host ::
RSS.