prof. Ing. Michal Haindl, DrSc.

The aim of this work is to analyze the perception of surface materials under variable illumination and observation conditions. The work will have the task of finding suitable static and dynamic visual stimuli and psychophysical verification of their relative importance for human perception and recognition of different materials.

The work is focused on the research of methods for recognizing and modeling the shape of bodies from video for virtual reality applications. Propose and implement a suitable method of automatic estimation of a 3D model from measured data using a video camera. Verify the method on selected models of sculptures and buildings.

Current advances in computed tomography scanning make it possible to virtually reconstruct Greek parchment scrolls burned 2,000 years ago after the eruption of Mount Vesuvius. The existing 280 scrolls found represent an enormous historical value. Their reading is a real revolution in modern methods of archeology and an international competition is being launched to solve it. The individual layers of the scroll must first be virtually unfolded, and the individual fragments connected to each other. Finding individual Greek letters on burnt scrolls is very difficult from CT scans, a possible solution is to use texture analysis methods. So far, only a few isolated words have been successfully read. The topic of the work is the improvement of the current state of reading burnt scrolls.

• Haindl, M. “Bidirectional Texture Function Modeling,” In: Handbook of Mathematical Models and Algorithms in Computer Vision and Imaging, chapter 28, pp. 1023–1064, ISBN 978-3-030-03009-4, DOI 10.1007/978-3-030-98661-2 103, Springer
• International Publishing, 2023.
• Mikes, S. - Haindl, M. “Texture Segmentation Benchmark,” IEEE Transactions on Pattern Analysis and Machine Intelligen-ce ,vol. 44, no. 9, pp. 5647-5663, DOI 10.1109/TPAMI.2021.3075916, ISSN 0162-8828, September, 2022.
• Haindl, M. - Havlicek, V. “Transfer Learning of Mixture Texture Models,” 12th Int. Conf. on Computational Collective Intelli-gence, ICCCI 2020, ISBN 978-3-030-63006-5, DOI 10.1007/978-3-030-63007-2 65, pp. 825–837, Lecture Notes in Artificial Intelligence, vol. 12496, 2020, Springer Nature Switzerland AG.

The thesis objective is to study environmental light field modelling methods from measured visual scenes applied to model BTF illumination in virtual reality application. Propose and implement an appropriate method for realistic illumination model estimated from real measured visual scenes. Verify this method on interior and exterior virtual reality models with variable illumination.

The thesis objective is to propose acquisition methods, capable to estimate advanced visual material representations (BTF, SVBRDF) directly from a real 3D object in natural lighting environment. Propose and implement an appropriate method for automatic visual surface material inference from hand held video camera sequences. Verify this method on the selected surface materials.

Design and implementation of a suitable statistical model that will allow modeling and prediction of the spread of viral infection between neigboring geographical areas. The model will also allow the evaluation of some selected epidemiological measures and their impact on the rate of spread of viral infection. The developed Markov model will be verified on COVID-19 coronavirus infection data.

The appearance of real surface materials significantly changes with variations in lighting and viewing parameters. Therefore, today's most advanced texture representations (BTF) require modeling reflectivity over a wide range of lighting parameters and camera locations using complex mathematical models. The aim of the work will be to develop and verify a new BTF model based on the theory of Markov random fields, which will improve the current state of physically correct modeling of material surfaces.

Design of suitable multispectral texture features for analysis of surface materials of visual scenes with variable observation conditions. The features will be derived from the multidimensional statistical models, their invariant modifications will be proposed, compared with the current best published alternative textural features, and applied to BTF data.

Malignant melanoma, as the most dangerous form of skin cancer, has been a rapidly growing threat over the past decades. The effective treatment requires its early diagnoses and surgical excision. The thesis objective is to find discriminative image features and a robust classifier which will allow to recognize skin cancer on colour skin images and to evaluate a treatment progress between several mutitemporal skin images.

The objective of this work is to critically evaluate the current state of unsupervised segmentation of image data and to develop an algorithm for unsupervised segmentation of dynamic color / multispectral / BTF images into individual homogeneous areas. The method will be based on the underlying multidimensional Markovian models and verified on the PTSB benchmark.

Verification of visual data modelling quality is a difficult and still unresolved problem due to the lack of existing mathematical criteria capable of approximating the human eye's perception. The work will aim at suggestion of an appropriate mathematical criterion which could quantify and rank simulation results according to their similarity to the measured visual template. The criterion will also be applicable to visual textures.

prof. Ing. Michal Haindl, DrSc.

doc. RNDr. Alena Šolcová, Ph.D.

Department of Applied Mathematics

Subsurface deposits of anthropogenic origin are often visible from aerial photographs in the form of changes in vegetation cover. These crop marks arise from different conditions for vegetation growth due to the changes in the local chemical composition of the soil caused by subsurface objects. This thesis's objective is to design and implement a method that is able to detect these crop marks in publicly available aerial images, provided by mapping services in the form of orthomaps. Such a method could significantly speed up the process of cataloging archaeological sites. The work offers a brief introduction to the concept of aerial archaeology and an overview of state-of-the-art methods of remote sensing and automatic detection with focus mainly on procedures where the number of positive examples is similarly limited as our dataset, constructed in this work. Three approaches to automatic detection are proposed, designed and implemented. The proposed methods are experimentally tested on the assembled dataset. The results are analyzed and the limitations of each approach are deduced. Both the first and second approach is based on detection of corners and linear segments and their spatial relations. The third approach, based on the template matching of crop marks of approximately known shapes, utilizing the generalized Hough transform algorithm yields the best results in crop mark recognition and is most promising for future research.

Thesis on DSpace

prof. Ing. Michal Haindl, DrSc.

doc. Ing. Štěpán Starosta, Ph.D.

Department of Applied Mathematics

The yellow-brown tortoise (Testudo hermanni) is on the list of endangered species, and it is therefore in the interest of animal rights activists to identify individuals of this species using biometric data from digital photographs. This work follows the work [3], which confirmed that biometric data can be used for unambiguous identification, however, the actual detection of plastrons was not very successful. Therefore, this work focuses mainly on a data set built from 323 low-quality photographs with different compositions of the scene, plastron orientation and resolution. Success rate achieved in the detection of the plastrons was 83.94 %.

Thesis on DSpace

prof. Ing. Michal Haindl, DrSc.

Department of Software Engineering

Dozens and possibly hundreds of textural features have already been introduced, but any comprehensive evaluation and comparison of the features is still lacking. We study and describe monospectral and multispectral features and based on this we create a general methodology for measuring the information quality of textural features. In this methodology, we classify features into categories, which creates a generalization layer that allows features to be evaluated generally and automatically. We will then incorporate this methodology in the creation of a multispectral textural benchmark with a web portal that allows experimentation with features. We will explain all phases of development from analysis, design of the user interface and its testing, to the actual implementation of the system. The created benchmark is made up of several components and can be expanded with other features, datasets for statistics, and last but not least, it is computationally scalable both vertically and horizontally.

Thesis on DSpace