Department of Software Engineering, Faculty of Nuclear Science and Physical Engineering, Czech Technical University in Prague

Supervisor: Adam Novozámský

Abstract: Human pose estimation and tracking are fundamental tasks in computer vision. The goal of this thesis is to provide a comparison between some available methods for pose estimation and tracking and create an application for their visualization. First, the problem of pose estimation is described along with the selected methods, selected datasets and their accuracy metrics. Next, pose tracking is similarly described with the accuracy metrics and the selected methods and datasets. The results for all methods on all datasets are then presented, discussed and compared. After that, the implementation and the usage of the application are described. At the end of the thesis, the goals are confronted and a conclusion is given which methods are the best in terms of accuracy and speed.

[download thesis]

Department of Mathematics, Faculty of Nuclear Science and Physical Engineering, Czech Technical University in Prague

Supervisor: Filip Sroubek

Supervisor: Zuzana Kadlecová - University of Cambridge

Abstract: This thesis delves into cellular endocytosis, focusing on clathrin-coated pits (CCPs) using advanced Total Internal Reflection Fluorescence - Structured Illumination Microscopy (TIRF-SIM). It presents a U-Net-based detection algorithm and a customized Multiple Hypothesis Tracking (MHT) algorithm for identifying and tracking CCPs in TIRF-SIM images. The performance of these computational tools is evaluated against human expert annotations, validating their effectiveness and identifying areas for improvement. This work contributes to the understanding of CCP behavior, blending high-resolution imaging with computational analysis, and has broader implications for cellular biology research.

[download thesis]

Department of Mathematics, Faculty of Nuclear Science and Physical Engineering, Czech Technical University in Prague

Supervisor: Filip Sroubek

Abstract: Super-resolution (SR) is a powerful imaging approach which offers significant gains in both lateral and axial resolution far beyond the diffraction limit of standard wide-field microscopy. This has lead to numerous breakthroughs in biology during the past 50 years of its existence. Among SR methods, structured illumination microscopy (SIM) has an major advantage with its flexibility of implementation and photon-efficiency due to which it plays a lead role in high-frequency in vivo cell acquisitions. First the elementary physics necessary for determining the transfer of light through an microscope is given by the derivation of the electro-magnetic (EM) field wave equation along with two of its most significant solutions for the application of microscopy optics. Next an condensed but informative overview of available SR methods is presented with the focus on comparing their utility in biological sensing applications. Intuitive conception behind the functioning of SIM is illustrated before the mathematical presentation of the fundamentals of harmonic SIM. An enumeration of examples where SIM is used in combination with other microscopy modalities and in various implementations demonstrates the flexibility the technique. Then an analysis of micro-beads images from a harmonic SIM acquisition is performed with the goal of enhancing the reconstruction. Two different methods of reconstruction are derived, implemented and demonstrated in multiple variations. The ability of limiting the number of acquisitions necessary for the reconstruction is discussed and shown alongside a commentary of further aspects, benefits and drawbacks of the reconstruction techniques. Further enhancements to the reconstruction techniques in the reduction of artefacts and improvement of visual quality and resolution is achieved by performing them with an estimated transfer function of the optical system. The estimation is performed using various procedures utilizing the model of the source structure in the object space of the determined by a custom novel algorithm inspired by multiple digital image processing algorithms. The fidelity over the SIM imaging technique currently used for the reconstruction of endocytosis at the cell membrane surface is attained by the data-driven transfer function estimate.

[download thesis]

Department of Mathematics, Faculty of Nuclear Science and Physical Engineering, Czech Technical University in Prague

Supervisor: Filip Sroubek

Abstract: This bachelor project is focused on application of deconvolution in astronomical imaging, with an emphasis on classical methods. We implement and optimize various deconvolution algorithms, which are subsequently tested on artificially blurred images as well as on real data. We also delve into the issues of sharpness measurement and artifact creation during deconvolution. We implement our own metric for determining the presence of artifacts in the image and for their quantification. The results of our work demonstrate how deconvolution algorithms can significantly improve the quality of astronomical images, although the quality of reconstruction is highly dependent on parameter settings and impulse response estimation. We also highlight the importance of sharpness measurement and artifact detection for the success of the deconvolution process.

[download thesis]

Department of Mathematics, Faculty of Nuclear Science and Physical Engineering, Czech Technical University in Prague

Supervisor: Adam Novozámský

Abstract: The subject of this work is gesture segmentation using classical image processing methods. Namely, thresholding, morphological transformations, histogram backprojection and others. Subsequently, an approach to gesture classification based on its contour and using Fourier descriptors is described. Finally, a comparison is made between the success rates of gesture classification using our proposed method and those based on a neural network model.

[dataset] [download thesis]

Department of Software Engineering, Faculty of Nuclear Science and Physical Engineering, Czech Technical University in Prague

Supervisor: Adam Novozámský

Abstract: Optical character recognition from image data is a demanding task in today’s world, because it is already impossible for humans to process a large amount of image data without automation. The Vienna Library owns over 350,000 digitized historical posters, from which the displayed text must be extracted. The aim of this work is to map existing text recognition methods and test them on selected datasets.

[download thesis]

Department of Mathematics, Faculty of Nuclear Science and Physical Engineering, Czech Technical University in Prague

Supervisor: Filip Sroubek

Abstract: The topic of this thesis is adaptive (dynamic) sampling and reconstruction of images. The objective is to compare five existing sampling methods (regular sampling, random sampling, modified Gaussian probability distribution, SLADS, PADIS) using two reconstruction methods ('griddata' and FSR). The first part of the thesis consists of research and a description of sampling methods and reconstruction methods used for comparison. The second part contains the results and discussion. For experiments we use MATLAB and Python, and four different datasets of 7 images were created for our experiments. Comparison is accomplished by using mean values of PSNR. Research showed that the adaptive sampling method PADIS overcomes other sampling methods for both reconstruction methods, starting from low sampling density.

[download thesis]

Department of Mathematics, Faculty of Nuclear Science and Physical Engineering, Czech Technical University in Prague

Supervisor: Adam Novozámský

Abstract: This thesis deals with face detection under unconstrained conditions. It aims to describe detection methods and select the appropriate approaches as well as datasets for testing them. The testing is done with the help of two publicly available collections (WFLW, CelebA) and one set created for the purposes of this thesis by labeling data from the Vienna City Library. The performance of these algorithms in less ideal conditions was also tested on more challenging subsets of the CelebA and WFLW datasets.

[download thesis]

Department of Algebra, Faculty of Mathematics and Physics, Charles University

Supervisor: Adam Novozámský

Abstract: The Wienbibliothek im Rathaus, Vienna City Library, collected over 300 thousand posters scanned in high quality from the last 100 years. Browsing and searching in such a large dataset is beyond human power. Therefore, a project was set up in cooperation with the Technical University of Vienna to test the possibilities of automatic data annotation on a selected sample. One of the requirements was Content-based Image Retrieval - retrieving images based on their visual content. This thesis reviews these techniques that emerged over the last decades. We focus on simple techniques based on colour, texture, and shape, as well as more advanced algorithms using convolutional neural networks. We implement these methods and compare their retrieval effectiveness on particular image datasets. Finally, we describe the functionality of a developed web application.

[download thesis]

Department of Software and Computer Science Education, Faculty of Mathematics and Physics, Charles University

Supervisor: Adam Novozámský

Abstract: At the end of every automotive assembly line, there is a quality control process where factory workers check produced cars for potential defects. The computer vision field, especially neural networks for images, have great potential to complement human staff in order to produce as safe and reliable cars as possible. In this thesis we focus on the validation, whether all four wheels on a single car match in size and type. We introduce and experiment with both neural networks and traditional computer vision techniques. The approach we use is to first detect the car then classify its wheels and try to estimate their size. In the end we build a functional prototype of the system that is running in real-time. The data for this thesis were recorded in Škoda Auto factory in Mladá Boleslav in cooperation with the company.

[download thesis]

Department of Algebra, Faculty of Mathematics and Physics, Charles University

Supervisor: Adam Novozámský

Abstract: The bachelor thesis deals with the design and description of a tool that can simplify the process of monitoring varroosis. Using image processing methods, it detects individual mites from the picture of the bottom board in the beehive. For better image resolution we used several images from one bottom board. Those images were then stitched into one image using image registration. In the first part we focused on the use of classical image processing methods that can detect varroa only on the basis of the approximation of the body of the mite with a parametrically descriptive curve. In the second part we used a more powerful convolutional neural network. During each of the 500 training epochs we saved the parameters of success. Finally, we inserted a test data into the trained network and compared it with expected outputs. The thesis contains a theoretical description of algorithms and methods, their use in our detector, and interpretation of results.

[download thesis]

Department of Mathematics, Faculty of Nuclear Science and Physical Engineering, Czech Technical University in Prague

Supervisor: Adam Novozámský

Abstract: Hand pose estimation plays a fundamental role in human computer interactions, moreover it allows us to analyze human behavior. The problem is nontrivial due to complicated hand variations caused by complex articulations, self-occlusions or shape, size and color ambiguities. We provide complex proposal of different detection and pose estimation methods and their evaluation. The comparison of deep map and RGB based pose estimation is provided and applied to real-life data from the driving simulator. Furthermore we design an annotation tool for RGB-D data which we used to produce a test dataset.

[download thesis]

Department of Mathematics, Faculty of Nuclear Science and Physical Engineering, Czech Technical University in Prague

Supervisor: Adam Novozámský

Abstract: The application of machine learning, and of its subset deep learning, has improved and enabled new techniques in the domain of computer vision. Sentinel-2 mission is specifically designed to provide high-resolution optical satellite imagery, which is suitable for monitoring surface vegetation. In addition to satellite images, we are provided with agricultural field data in 2019 by State Agricultural Intervention Fund. Combining these two data sources, datasets for field crop classification into thirteen classes are prepared. Both feature-based classifiers and convolutional neural networks are then applied on these pre-processed datasets and the results of each classifier and its performance are analyzed, discussed and compared.

[download thesis]

Department of Mathematics, Faculty of Nuclear Science and Physical Engineering, Czech Technical University in Prague

Supervisor: Adam Novozámský

Abstract: In most automobile factories, the quality inspection process is mainly based on vision control, which is often insufficient and unstable. The artificial intelligence can improve some parts of this quality process. The aim of our project was to develop the "Proof-of-concept" of such an automated visual inspection system. We focused on the level of cooling liquid. The analysis includes data collection and labeling, pre-processing, feature extraction and classification.

[download thesis]