|
|
Radial Basis Function Neural Network
Nevoral, Leoš ; Rozman, Jaroslav (referee) ; Zbořil, František (advisor)
The thesis focuses on explaining the field of RBF neural networks, specifically RCE networks, through a demo application. The demo application primarily visualizes the network learning process, as well as the state of the neural network and different shapes of basis functions. Additionally, the utilization of EBF neural networks is explored. Conventional approaches to EBF networks are compared and tested against new design of OEBF network. Which is based on deriving the elliptical areas from euclidean distance from both focal points of ellipse. The new design shows no signs of improving the properties of these networks and rather produces results almost identical to those of the classic RCE network, which are, however, several percentage points less accurate. Finally, methods for improving this solution in future are proposed.
|
|
|
Surface defect detection of metal parts based on neural networks
Hadwiger, Tomáš ; Jonák, Martin (referee) ; Ježek, Štěpán (advisor)
The goal of this thesis is focused on surface anomaly detection on metal parts. The goal was to implement different neural network architectures using the method CutPaste and compare them on three different datasets: MVTec AD, MPDD, MPDD2. For the object classes of the dataset MVTec AD the most accurate architecture turned out to be ResNet-18 with average precision of 84,45 AUROC, for the materials it was the EfficientNet architecture with average precision of 87,22 AUROC. For the MPDD and MPDD2 datasets, the most accurate architecture was ResNet50 with average precision of 88,64 AUROC and 61,10 AUROC respectively. Based on the measure values, the most difficult dataset for anomaly detection turned out to be MPDD2.
|
|
|
Automated Metadata Extraction From Document Images
Křivánek, Jakub ; Vaško, Marek (referee) ; Kohút, Jan (advisor)
This Bachelor thesis addresses the problem of extracting structured data from scans of documents from Czech libraries. The aim of the thesis is to simplify the time-consuming manual process for librarians. I focused on creating datasets from documents of Czech libraries and on detecting metadata on these datasets. I created one dataset for books and another for periodicals. Detection was performed by classifying lines read from the documents. This utilized a fully connected neural network and a network employing a Transformer Encoder. The second method of metadata detection is based on object detection in document scans using the YOLOv8 model. Detection using the fully connected neural network achieves an F1 score of 0.83 on the book dataset and 0.78 on the periodicals dataset. The Transformer Encoder network achieves F1 scores of 0.84 on the book dataset and 0.59 on the periodicals dataset. The YOLO model achieves an F1 score of 0.86 (confidence at 0.549) on the book dataset and 0.7 (confidence at 0.336) on the periodicals dataset.
|
|
|
Adaptivní systém pro řízení osvětlení ve Smart Home
Valík, Tomáš ; Rozman, Jaroslav (referee) ; Janoušek, Vladimír (advisor)
The thesis deals with the issue of lighting control in smart homes. In most smart homes, it is necessary to manually control the lighting using switches or mobile devices. The thesis introduces an adaptive control system based on recurrent neural networks, which gradually learns user manipulation with the lighting and eventually begins to independently control the lighting after a certain period of time.
|
|
|
Application for Automatic Evaluation of the Fidelity of the Generated Facial Image
Šotola, Jiří ; Semerád, Lukáš (referee) ; Goldmann, Tomáš (advisor)
This work focuses on the design and implementation of an application for verifying the fidelity of a synthetically generated images, which, due to the vastness of this topic, is aimed at verifying the similarity of the facial features of the original image and the image generated from it. For this application, a Gen_Verifier model is developed based on Siamese networks, which uses the contrastive loss. This model was trained and tested on the LFW dataset, where it reached an accuracy of 91 %. The StarGAN model is used to test the generated images, which generated facial images with changes in hair color, gender and age. The resulting testing on the generated images showed that the StarGAN model produces faces that are similar in 87.53 % cases.
|
|
|
Vision Transformers for Facial Recognition
Strýček, Šimon ; Kišš, Martin (referee) ; Špaňhel, Jakub (advisor)
This thesis focuses on applying vision transformer-based neural networks to face recognition related tasks. It focuses on exploring modern vision transformer (ViT) architectures, experimenting with alternative data, and finding the suitable parameters to train ViTs to compete with the already established dominance of convolutional neural networks in face recognition. The goal of this work was to show the suitability of vision-transformers for face recognition. The output of this work contains results of various experiments, demonstrations of benefits and drawbacks of some of the modern and popular ViTs, the definition of an optimal setup when wanting to employ vision transformers for facial recognition, and interesting observations from working with vision transformers.
|
|
|
Anthropometric Evaluation of Generated Face Images
Mikyšek, Jakub ; Rydlo, Štěpán (referee) ; Goldmann, Tomáš (advisor)
This work is focused on comparing artificially generated faces with real images by analyzing key facial landmarks and measuring the proportions between these landmarks. It also explores areas related to artificial neural networks, focusing on GANs that can generate artificially generated facial images. It explores their process, architecture and available models. The aim of the work is to evaluate how artificially generated images differ from real ones, and to find out in which proportions the difference is the largest.
|
|
|
Web application integrating artificial intelligence techniques into the correlation rule creation process
Šibor, Martin ; Caha, Tomáš (referee) ; Safonov, Yehor (advisor)
Currently, as digitalization becomes an integral part of all areas of our lives, the complexity and sophistication of cyber threats are constantly increasing. A key element in the fight against these cyber threats is security monitoring. An important tool for security monitoring are SIEM systems, which allow for early detection and response to potential attacks based on correlation rules. The main contribution of this work is the design and implementation of a web application that integrates artificial intelligence techniques into the process of creating and managing correlation rules for security monitoring systems, with the aim of streamlining the process of creating, modifying, and understanding correlation rules. The work first provides a theoretical introduction to the field of natural language processing and modern neural networks, particularly the transformer architecture, which is the basis of generative artificial intelligence models (e.g., ChatGPT, Gemini). It then introduces the principles of security monitoring, log management systems, the concept of correlation rule generalization, and, last but not least, the challenges associated with managing and maintaining correlation rules, which the integration of artificial intelligence into these processes significantly reduces. The practical part of the work describes the design and implementation of a web application that utilizes the gpt-4 and gpt-3.5-turbo models from OpenAI and the Gemini Ultra 1.0 model from Google for creating new correlation rules, modifying existing rules, and explaining and interpreting them for easier understanding and faster deployment. The application is designed with user-friendliness and efficiency in mind. The results of the work show that the integration of artificial intelligence into the correlation rule creation process brings significant efficiency improvements. The web application allows users to easily create and modify correlation rules. The application also allows users to better understand correlation rules, enabling them to respond to potential threats more quickly.
|
|
|
Automatic Transcription of Air-Traffic Communication to Text
Nevařilová, Veronika ; Veselý, Karel (referee) ; Szőke, Igor (advisor)
This thesis focuses on fine-tuning Whisper, an automatic speech recognition model developed by OpenAI, on Czech and English recordings of air-traffic communication. It provides a fundamental insight into automatic speech recognition, neural networks and transformer architecture. Further, data collection and annotation is also described and after that it details the process and outcomes of Whisper’s training on two different transcription formats – full, where the model learns to transcribe recordings word by word, and abbreviated, which is more suitable for quick navigation and more natural for air traffic controllers.
|
|
|
Tracking people based on their clothing in multi-camera systems
Sivak, Mykyta ; Přinosil, Jiří (referee) ; Číka, Petr (advisor)
This bachelor thesis focuses on the development and implementation of an algorithm for tracking individuals in multi-camera systems based on clothing pattern analysis. The aim was to design a system capable of tracking an individual in various positions and frames, using the Region of Interest (RoI) technique. The study begins with a comprehensive review of the existing literature on object tracking in video sequences, with a special focus on RoI tracking techniques. During the research, a new algorithm was developed and implemented that utilizes clothing patterns as the primary identification element for tracking and re-identifying individuals across different camera shots. The algorithm was experimentally validated on datasets containing video sequences from various environments, allowing for a detailed analysis of its effectiveness and reliability. The experimental results demonstrate that the proposed system achieves significant accuracy and efficiency compared to traditional methods and is particularly effective in challenging situations where other methods fail. The thesis concludes with an evaluation of the conducted experiments along with recommendations for future extensions and improvements of the system. Potential challenges and ethical aspects, including issues of privacy and personal data processing, are also discussed.
|
guest ::
