|
|
A Reduced Neural Network for Classifying the Presence of People in an Image
Stanček, Rastislav ; Rydlo, Štěpán (oponent) ; Goldmann, Tomáš (vedoucí práce)
This thesis focuses on the topic of computer vision, more specifically, on classifying people's presence in image data. The goal is to create a reduced neural network utilizing knowledge distillation. Object classification and detection is a computationally an expensive operation. A student model created utilizing knowledge distillation shows equivalent accuracy while being smaller and having better inferencing speed compared to the teacher model. Such model can be interdisciplinarily utilized on end devices having relatively low computational capabilities.
|
|
| |
|
|
Implementing gesture recognition on ARM as an alternative to traditional device control
Gajdošík, Richard ; Zbořil, František (oponent) ; Kočí, Radek (vedoucí práce)
This bachelor's thesis focuses on the development and implementation of a gesture recognition system on ARM architecture, utilizing the i.MX 93 board and TensorFlow Lite. The project is grounded in the application of neural networks for the recognition of hand gestures, offering an alternative to traditional device control methods. An integral part of the work involves a comprehensive analysis of existing gesture recognition solutions, identifying their strengths and potential improvements. The thesis elaborates on the design, development, and optimization of a real-time gesture recognition model specifically for ARM chips, emphasizing efficiency and performance. Additionally, the thesis covers the creation of a demonstrative application that visually represents recognized gestures. User testing is conducted to evaluate the practicality and user experience of the gesture recognition system, providing valuable feedback for future enhancements.
|
|
|
Neural Networks for Video Quality Enhancement
Sirovatka, Matej ; Juránek, Roman (oponent) ; Hradiš, Michal (vedoucí práce)
In this thesis, a new method for video super-resolution is proposed. The method is based on the idea of using deformable convolutional layers together with optical flow to align features from multiple sequential video frames. This novel module is then used in a U-Net-like deep neural network to predict high-resolution frames. The proposed method is evaluated on a dataset containing real-life scenes and compared to other methods. Multiple different configurations of the proposed method are tested and the results are analyzed. The results of the experiments show promising results, with the model outperforming bilinear interpolation, and single-frame methods. Multiple different architectures of the feature alignment module together with the rest of the U-Net architecture are tested, showing that using Vgg19 as the encoder of the U-Net gives the best results.
|
|
|
Creating a Python-based Automated System for Recognizing Emotions from Facial Expressions.
Zima, Samuel ; Malik, Aamir Saeed (oponent) ; Hussain, Yasir (vedoucí práce)
This thesis examines facial expression recognition (FER) using deep learning by focusing on its application in devices with limited memory and computational resources. It begins by researching emotions and facial expressions from psychological, biological, and sociological perspectives. The core of this thesis involves the design and implementation of an automated FER system using the FER-2013 dataset. This system uses a customized SqueezeNet architecture enhanced with a simple bypass, dropout layers and batch normalization layers. This system achieves an accuracy of 66.37 % on the FER-2013 dataset. For comparative analysis, this model was compared with a customized VGG16 architecture which achieved an accuracy of 65.09 %. This thesis provides valuable insights into the development of smaller, more efficient machine learning models for FER which are usable in a wide range of devices, including low-performance CPUs and embedded devices.
|
|
|
Automated compression of neural network weights
Lorinc, Marián ; Sekanina, Lukáš (oponent) ; Mrázek, Vojtěch (vedoucí práce)
Convolutional Neural Networks (CNNs) have revolutionised computer vision field since their introduction. By replacing weights with convolution filters containing trainable weights, CNNs significantly reduced memory usage. However, this reduction came at the cost of increased computational resource requirements, as convolution operations are more computation intensive. Despite this, memory usage remains more energy-intensive than computation. This thesis explores whether it is possible to avoid loading weights from memory and instead functionally calculate them, thereby saving energy. To test this hypothesis, a novel weight compression algorithm was developed using Cartesian Genetic Programming. This algorithm searches for the most optimal weight compression function, aiming to enhance energy efficiency without compromising the functionality of the neural network. Experiments conducted on the LeNet-5 and MobileNetV2 architectures demonstrated that the algorithm could effectively reduce energy consumption while maintaining high model accuracy. The results showed that certain layers could benefit from weight computation, validating the potential for energy-efficient neural network implementations.
|
|
|
Deep Learning for 3D Mesh Registration
Pukanec, Dávid ; Beran, Vítězslav (oponent) ; Španěl, Michal (vedoucí práce)
The problem of mesh alignment is often solved through point cloud registration. Numer- ous deep learning-based registration methods are published every year achieving state-of- the-art results. Based on their core concepts, the methods can loosely be divided into correspondence-based and correspondence-free. Even though comparisons of individual methods exist, the cross-evaluations of both categories are lacking. In this work, a deeper evaluation of Lepard and FINet models is presented. For this purpose, the ModelNet40 and Teeth3DS datasets are used. The experiments show that FINet is able to align unseen shapes, obscured by partiality and noise with a translation error of 4.16% of model size and a rotation error of 3.640 degrees. While Lepard manages this with a translation error of 6.73% of model size and a rotation error of 7.265 degrees.
|
|
|
Detection of Nudity in an Image Data
Pešková, Daniela ; Orság, Filip (oponent) ; Goldmann, Tomáš (vedoucí práce)
The focus of this thesis is the creation of a tool capable of detecting nudity in image data. This is achieved by training a model to detect incriminated body parts and creating an algorithm capable of detecting skin. The resulting tools can be used for automatic nudity detection in images. The first part of the thesis focuses on the theory of neural networks and computer vision, with an emphasis on skin detection. The second part discusses the approach chosen for creating the dataset, the process of creation and training the model capable of detecting nudity in images, as well as the algorithmic approach.
|
|
|
Guided Reinforcement Learning for Motor Skills
Karabelly, Jozef ; Herout, Adam (oponent) ; Hradiš, Michal (vedoucí práce)
This thesis aims to present an overview of the current state of research in guided reinforcement learning for motor skills and identify potential research paths. Besides, the thesis introduces an improved method for learning physically simulated character animations based on the current techniques. The pre-trained model shows the ability to perform well on various new tasks. A custom dataset was collected explicitly for pre-training the model introduced in this thesis. Future improvements and possible research paths are proposed based on the experiments' results.
|
|
|
AI-based classification of RF signals
Turák, Samuel ; Ulovec, Karel (oponent) ; Polák, Ladislav (vedoucí práce)
This thesis focuses on Deep learning-based radio frequency (RF) signal classification. For this purpose, three neural networks are selected and introduced: Convolutional Neural Network (CNN), Gated Recurrent Unit Network (GRU) and Convolutional Gated Deep Neural Network (CGDNN). All are trained and evaluated on multiple datasets, influenced by different RF impairments, for wireless standard classification. The waveforms in these datasets have been created by the Wireless Waveform Generator app in MATLAB. One publicly available modulation classification dataset is also being tested on the models. The performed approaches of data preprocessing, model training and model evaluation are implemented in the programming environment Python, utilizing libraries such as Scikit-learn and Keras. The obtained results are evaluated and discussed.
|
host ::
RSS.