|
|
Measuring the thickness of contamination layers in scanning electron microscopy using image processing
Macek, Matěj ; Munzar, Milan (oponent) ; Čadík, Martin (vedoucí práce)
The motivation for this thesis arises from the aim of the Thermo Fisher Scientific company to develop a method to measure the thickness of contamination layers in scanning electron microscopy (SEM) images through advanced image processing techniques. The primary aim is to create automated methods for quantifying contamination in images that adversely impact imaging in material science research, using image processing techniques. In this study, we collect a dataset of images with contamination and manually annotate masks for each image. These annotations will serve to fine-tune and evaluate the effectiveness of the methods we propose. By employing a combination of edge detection algorithms and machine learning models, specifically a fine-tuned DeepLabv3 network, this work enhances the precision and efficiency for contamination detection. The edge Detection-Based Contamination Analyzer (EDCA) utilizes traditional image processing methods, while the DeepLabv3 model introduces a machine learning approach to robustly handle diverse imaging conditions. Comparative analyses demonstrate the effectiveness of these methods in providing reliable, scalable, and detailed measurements of contamination layers, significantly contributing to the field of materials science.
|
|
|
Simulation of Biological Processes Using Asynchronous Cellular Automata and Machine Learning
Kališ, Vojtěch ; Bidlo, Michal (oponent) ; Fritz, Karel (vedoucí práce)
This thesis explores the fusion of asynchronous cellular automata and machine learning techniques for simulating complex biological processes. Its main focus is on showcasing the inherent potential of a computational framework constructed through combining the parallelism of an asynchronous cellular automata updating model with the predictive capabilities of machine learning algorithms. This study aims to demonstrate the qualities of such hybrid approach by implementing three mathematical cellular automata models of increasing complexity—that is, listed based on their level of complexity, Conway’s Game of Life, SmoothLife and Lenia—in their basic form and then integrate machine learning into the function of the latter two, comparing the results of both approaches afterwards.
|
|
|
Anion-exchange enabled tuning of caesium lead mixed-halide perovskites for high-energy radiation detection
Matula, Radovan ; Friák, Martin (oponent) ; Dvořák, Petr (vedoucí práce)
Lead halide perovskites (LHPs) with their unprecedented functional qualities which are only enhanced by the simple band gap tuning, have taken the world of semiconductors by storm. The process of anion exchange, possible even post-synthesis, allows for band gap tuning of LHPs, resulting in lead mixed-halide perovskites (LMHPs), thus expanding their potential for applications, notably in tuneable detectors. The widespread adoption of LMHPs is, however, hindered by their chemical instability, which leads to halide segregation in the material, seriously inhibiting reliable operation of any LMHP-based device. Understanding the kinetics of the halide segregation over extended periods remains a challenge, motivating the use of theoretical simulations like Monte Carlo (MC) methods. Yet, MC simulations rely on well-defined potential energy surfaces (PES), typically derived from computationally intensive density functional theory (DFT) calculations. In this thesis, we propose a novel approach for constructing well-defined PES from high-fidelity DFT data with fraction of the computational load. Utilizing activation-relaxation technique noveau (ARTn) motivated searches for transition points in the PES combined with state-of-the-art machine learning approaches, we aim to to significantly reduce computational costs. Additionally, employing classical theory, we assess the detection capabilities of selected LMHPs.
|
|
|
Automatický přepis řeči s podporou code switching
Bílek, Štěpán ; Karafiát, Martin (oponent) ; Szőke, Igor (vedoucí práce)
Tato práce se zabývá problematikou automatického rozpoznávání řeči. Zaměřuje se na rozpoznávání audia obsahující vícejazyčné promluvy, tzv. code-switching. Problém nedostatku vícejazyčných dat pro trénování je řešen kombinováním nahrávek v angličtině a němčině dohromady. Pro co největší přiblížení ke skutečné dvojjazyčné řeči je část datasetů tvořena spojováním nahrávek podobných mluvčích. Na vytvořených datech je trénován a testován model Whisper. Ten v původní neadaptované verzi dosahuje chybovosti až 70 %. Nejlepší modely trénované na kombinovaných datasetech dosahují chybovosti jen lehce přes 7 %. Výsledky této práce ukazují způsoby jak modely trénovat, aby dosahovaly co nejlepších výsledků.
|
|
|
Machine learning in audio effects
Sychra, Jakub ; Mošner, Ladislav (oponent) ; Černocký, Jan (vedoucí práce)
Reverse engineering audio effects from mixed tracks is a complex topic requiring signal processing and music engineering experience. This work aims at creation of a system capable of identifying the sequence and parameters of guitar effects from a mixed audio track. Training data was created using clean guitar sounds from IDMT-SMT-Audio-Effects, augmented by known effects (BitCrush, Chorus, Clipping, Compressor, Delay, Distortion, High-pass filter, Ladder filter, Low-pass filter, Limiter, Phaser and Reverb), all implemented with a Python wrapper around standard VST effects. The system is based on VGGish neural network architecture with several classification (presence of effects) and regression (parameters of effects) heads. The performance of the algorithm is evaluated on classification and regression accuracy, as well as in informal listening tests.
|
|
|
Meta-heuristické algoritmy pro výběr příznaků v klasifikaci srdečních onemocnění
Švestková, Tereza ; Odstrčilík, Jan (oponent) ; Mézl, Martin (vedoucí práce)
Tato práce je věnována postupům při výběru relevantních příznaků pro klasifikační úlohy týkající se srdečních onemocnění. Optimální výběr příznaků je klíčovým faktorem pro správnou funkčnost klasifikačních modelů a v případě medicíny pro zlepšení diagnostiky. Teoretická část pojednává o obecné klasifikační úloze ve strojovém učení. Dále jsou podrobněji popsány některé klasické postupy i novější meta-heuristické algoritmy pro efektivní selekci příznaků. Praktická část se věnuje aplikaci některých popsaných algoritmů na datové soubory související se srdečním onemocněním. Na základě ověření validity výsledku klasifikačního modelu podle vybraných příznaků běžnými postupy i evolučními algoritmy jsou diskutovány výhody a přínosy upřednostnění meta-heuristických algoritmů.
|
|
|
Detekce malware domén pomocí metod strojového učení
Ebert, Tomáš ; Poliakov, Daniel (oponent) ; Hranický, Radek (vedoucí práce)
Tato bakalářská práce se zabývá detekcí malware domén pomocí metod strojového učení na základě různých informací získaných o doméně (DNS záznamy, geolokační údaje atd.). S rychle rozšiřujícími se hrozbami, nejen formou malwaru, jsou často současné přístupy nedostačující ať už jen rychlostí detekce malware domén, nebo celkovým rozeznáním, zda se jedná o nebezpečnou doménu. Výstupem této práce je natrénovaný model klasifikátoru XGBoost, jehož výhodou je rychlá a efektivní detekce v reálném čase oproti detekci pomocí černých listin, které získávají data domén často s týdenním zpožděním. Pro tento model bylo získáno 131 tisíc malware domén, pomocí kterých bylo možné získat model s vysokými hodnotami. Pomocí experimentů bylo dosaženo skóre F1 96.8786 % u klasifikátoru XGBoost s poměrem falešně pozitivních detekcí 0.004887.
|
|
|
Optimization of Classification Models for Malicious Domain Detection
Pouč, Petr ; Jeřábek, Kamil (oponent) ; Hranický, Radek (vedoucí práce)
This thesis focuses on the development of advanced methods for malicious domain name detection using optimization techniques in machine learning. The thesis investigates and evaluates the effectiveness of different optimization strategies for classification. As evaluation tools, I selected classification algorithms that differ in their approach, including deep learning, decision tree techniques, or hyperplane search. These methods are investigated in terms of their ability to effectively classify domain names depending on the implemented optimization techniques. Optimization strategies include the creation of ground-truth datasets, application of data processing methods, advanced feature selection, solving the class imbalance problem, and hyperparameter tuning. The final part of the paper presents a detailed analysis of the benefits of each optimization approach. The experimental part of the study demonstrates exceptional results by combining several methodologies. The top CNN models obtained up to 0.9926 F1 while lowering FPR to 0.3%. The contribution of this study is to provide specific methodologies and tactics for the successful identification of malicious domain names in the cybersecurity area.
|
|
|
Analyzing a person’s handwriting for recognizing his/her emotional state
Chudárek, Aleš ; Matoušek, Jiří (oponent) ; Malik, Aamir Saeed (vedoucí práce)
Emotion recognition from handwriting is a challenging and interdisciplinary task that can provide insights into the psychological and emotional aspects of the writer. In this study, we developed and evaluated a machine learning model that can predict the emotional state of a writer from their handwriting samples. We utilized the EMOTHAW dataset, which consists of handwriting and drawing samples from subjects whose emotional states are measured by the DASS test, which gives a score for depression, anxiety, and stress and the CIU Handwritten database for verification and experimentation. We extracted a large number of features that are inspired by the standard graphology work, as well as features that are specific to online data. We used ANOVA to select statistically significant features and normalized the data using Z-Score, MinMax, IQR or Log. We reduced the dimensionality of the features using Principal Component Analysis (PCA) and Linear Discriminant Analysis (LDA). We employed a meta approach Ensemble learning that seeks to reduce the errors of a single model by exploiting the diversity and complementarity of multiple models. The structure of our classifier is dependent on multiple arguments resulting in over 300,000 different configurations. We optimized arguments using argument freezing. We found the best classifiers for binary and trinary classification for each emotion, resulting in six optimal models. We evaluated our models using different metrics, such as accuracy, precision, recall, and F1-score. Our models reached adequate results in all metrics. In addition to finding the classifiers, this thesis explored the importance of each extracted feature, providing a sorted list of the most significant features used for emotion recognition from handwriting. We also enhanced the EMOTHAW database by identifying tasks that are more indicative of specific emotions, thereby reducing the need for a full task battery for emotional analysis.
|
|
|
Game with Procedural Enemies
Teplan, Kristof ; Pečiva, Jan (oponent) ; Milet, Tomáš (vedoucí práce)
Procedural content generation has gained significant popularity, particularly in the video game industry. However, the area of procedural enemy generation in games remains relatively unexplored. This thesis focuses on the development of a rogue-like 2D game where the player battles against procedurally generated robots. The game is implemented in Unity Engine, utilizing a custom matrix L-system to generate robots from blocks with various functionalities. The artificial intelligence of the robots is powered by a neural network trained using reinforcement learning with the Unity ML-Agents toolkit. All systems are integrated into a final game that captures the classic features of the rogue-like genre.
|
host ::
RSS.