|
|
Forensic method for recognizing the authenticity of artworks using multispectral analysis
Lánský, David ; Mezina, Anzhelika (referee) ; Burget, Radim (advisor)
Detecting forgeries is crucial for protecting the art market and preserving the authenticity of artworks. This thesis focuses on forgery detection using convolutional neural networks (CNNs). The main goal was to develop advanced methods capable of identifying anomalies, and thus potential forgeries, in images with their X-ray photographs. During this research, U-net architectures and binary semantic segmentation techniques were applied, enabling successful anomaly detection. The main contribution of this work is 112 models of four different U-net and U-net++ architectures, which effectively highlight anomalies through the method of binary semantic segmentation. The models were trained on a set of images with their synthetically created X-ray images and artificially generated anomalies. In this way, the models can detect lead spots, nails, layers of hidden paintings, and other defects, while also being able to ignore insignificant elements, such as picture frames and overexposed X-ray images. The testing of the models occurred in two phases. In the first phase, they were evaluated using the IoU metric on a set of 400 synthetically generated data, where in the best cases, they achieved up to 83.5 % IoU. In the second phase, they were evaluated subjectively on images with real X-rays and natural anomalies. This approach combines traditional X-ray techniques with modern computer vision, revealing deviations that might be overlooked during standard visual inspection. By bridging these technologies, this work opens new possibilities for the protection of art collections and provides a solid foundation for further research in the field of art forgery detection using artificial intelligence.
|
|
|
Generative Neural Network for Creating Synthetic Photorealistic Images
Hora, Adam ; Přinosil, Jiří (referee) ; Říha, Kamil (advisor)
The main objective of this work is to select and design a neural network model that will be able to generate realistic images thematically fitting the selected dataset. The architecture used for the solution is Deep convolutional generative adversarial network. This network is than implemented in the Python programming language using the Tensorflow application programming interface and its included interface Keras. Finally, the model is trained on the selected dataset and the resulting generated images are presented. The final model and individual images are then evaluated using various quality assessment methods.
|
|
|
Segmentation of Electrocardiographic Signals Using Deep Learning Methods
Hejč, Jakub ; Černý, Martin (referee) ; Halámek, Josef (referee) ; Kolářová, Jana (advisor)
The thesis deals with deep learning methods for the segmentation of surface and intracardiac electrocardiographic recording with focus on atrial activity. The theoretical part introduces current segmentation aproaches of electrocardiographic signals. Issues related to the development of deep learning models in context of standard ECG databases were also discussed. We proposed a pipeling for processing multimodal electrophysiology data from interventional procedures in order to build reliable training datasets. A deep model for segmentation of intracardiac recordings based on a modified residual architecture was proposed. A series of experiments was conducted to evaluate the effect of both model and dataset properties on segmentation quality. The annotation methodology of recordings with atrial fibrillation proved to be a crucial factor. Properties of loss function and type of data augmentation were revealed as secondary important parameters. A novel P wave segmentation method for incomplete references was proposed in the thesis. The approach was inspired by the deep contrast learning. It was modified to distinguish local segments of signals at different levels of abstraction of the extracted feature maps. Results were analyzed using standard quality metrics and post-hoc visual analysis. In some cases, a statistical comparison of experiments for different settings was performed. The results of the work showed that it is possible to use intracardiac signals for embedding a vector representation of local atrial activation into deep models.
|
|
|
Deep Learning for Image Stitching
Šilling, Petr ; Beran, Vítězslav (referee) ; Španěl, Michal (advisor)
Sešívání obrázků je klíčovou technikou pro rekonstrukci objemů biologických vzorků z překrývajících se snímků z elektronové mikroskopie (EM). Současné metody zpracování snímků z EM k sešívání zpravidla využívají ručně definované příznaky, produkované například technikou SIFT. Nedávný vývoj však ukazuje, že konvoluční neuronové sítě dokáží zlepšit přesnost sešívání tím, že se naučí diskriminativní příznaky přímo z trénovacích obrázků. S ohledem na potenciál konvolučních neuronových sítí tato práce navrhuje sešívací nástroj DEMIS, který staví na pozornostní síti LoFTR pro hledání shodných příznaků mezi páry obrázků. Dále práce navrhuje novou datovou sadu generovanou dělením obrázků z EM s vysokým rozlišením na pole překrývajících se dlaždic. Výsledná datová sada je použita pro dotrénování sítě LoFTR a k vyhodnocení nástroje DEMIS. Experimenty na dané datové sadě ukazují, že nástroj je schopen nalézt přesnější shody mezi příznaky než SIFT. Navazující experimenty na obrázcích s vysokým rozlišením a malými překryvy mezi dlaždicemi dále poukazují na výrazně vyšší robustnost oproti metodě SIFT. Dosažené výsledky celkově naznačují, že hluboké učení může vést k prospěšným změnám v oblasti EM, například k umožnění menších překryvů mezi snímanými obrázky.
|
|
|
Intracranial hemorrhage localization in axial slices of head CT images
Kopečný, Kryštof ; Chmelík, Jiří (referee) ; Nemček, Jakub (advisor)
This thesis is focused on detection of intracranial hemorrhage in CT images using both one-stage and two-stage object detectors based on convolutional neural networks. The fundamentals of intracranial hemorrhage pathology and CT imaging as well as essential insight into computer vision and object detection are listed in this work. The knowledge of these fields of studies is a starting point for the implemenation of hemorrhage detector. The use of open-source CT image datasets is also discussed. The final part of this thesis is a model evaluation on a test dataset and results examination.
|
|
|
Detection of cells in confocal microscopy images
Hubálek, Michal ; Štursa, Dominik (referee) ; Škrabánek, Pavel (advisor)
The goal of the thesis was to create an application that automatically detects healthy cardiomyocytes from images captured by a confocal microscope. The thesis was created based on the specific needs of researchers from the Slovak Academy of Sciences.The application will facilitate and increase the efficiency of their research,because until now they have to evaluate the images and search for suitable cells manually. The RetinaNet convolutional neural network is used for detection and has been implemented in a user-friendly desktop application. The application also automatically records and stores coordinates of detected cells which can be used for capturing cells in higher image quality. Another advantage of the developed application is its versatility, which allows to train detection on other data, making it applicable to other projects. The result of this work is a functional, standalone and intuitive application that is ready to be used by researchers.
|
|
|
A convolutional neural network for image segmentation
Mitrenga, Michal ; Petyovský, Petr (referee) ; Jirsík, Václav (advisor)
The aim of the bachelor thesis is to learn more about the problem of convolutional neural networks and to realize image segmentation. This theme includes the field of computer vision, which is used in systems of artificial intelligence. Special Attention is paid to the image segmentation process. Furthermore, the thesis deals with the basic principles of artificial neural networks, the structure of convolutional neural networks and especially with the description of individual semantic segmentation architectures. The chosen SegNet architecture is used in a practical application along with a pre-learned network. Part of the work is a database of CamVid images, which is used for training. For testing, a custom image database is created. Practical part is focused on CNN training and searching for unsuitable parameters for network learning using SW Matlab.
|
|
|
Vehicle Control via Reinforcement Learning
Maslowski, Petr ; Uhlíř, Václav (referee) ; Šůstek, Martin (advisor)
The goal of this thesis is a creation of an autonomous agent that can control a vehicle. The agent utilizes reinforcement learning that uses neural networks. The agent interprets images from the front vehicle camera and selects appropriate actions to control the vehicle. I designed and created reward functions and then experimented with hyperparameters setup. Trained agent simulate driving on the road. The result of this thesis shows a possible approach to control an autonomous vehicle agent using machine learning method in CARLA simulator.
|
|
|
Vehicle Counting in Still Image
Vágner, Filip ; Juránek, Roman (referee) ; Špaňhel, Jakub (advisor)
The goal of this work is to compare models of convolutional neural networks designed to count vehicles in a static image using density estimation with a focus on different sizes of objects in the scene. A total of four models were evaluated - Scale Pyramid Network, Scale-adaptive CNN, Multi-scale fusion network and CASA-Crowd. The evaluation was done on three data sets - TRANCOS, CARPK, PUCPR+. Scale Pyramid Network achieved the best results. The model reached 5.44 in the Mean Absolute Error metric and 9.95 in the GAME(3) metric on TRANCOS dataset.
|
|
|
Identification of vertebrae type in CT data by machine learning methods
Matoušková, Barbora ; Kolář, Radim (referee) ; Chmelík, Jiří (advisor)
Identification of vertebrae type by machine learning is an important task to facilitate the work of medical doctors. This task is embarrassed by many factors. First, a spinal CT imagining is usually performed on patiens with pathologies such as lesions, tumors, kyphosis, lordosis, scoliosis or patients with various implants that cause artifacts in the images. Furthermore, the neighboring vertebraes are very similar which also complicates this task. This paper deals with already segmented vertebrae classification into cervical, thoracic and lumbar groups. Support vector machines (SVM) and convolutional neural networks (CNN) AlexNet and VGG16 are used for classification. The results are compared in the conclusion.
|
guest ::
