|
Deployment of deep learning-based anomaly detection systems: challenges and solutions
Ježek, Štěpán ; Burget, Radim
Visual anomaly detection systems play an important role in various domains, including surveillance, industrial quality control, and medical imaging. However, the deployment of such systems presents significant challenges due to a wide range of possible scene setups with varying number of devices and high computational requirements of deep learning algorithms. This research paper investigates the challenges encountered during the deployment of visual anomaly detection systems for industrial applications and proposes solutions to address them effectively. We present a model use case scenario from real-world manufacturing quality control and propose an efficient distributed system for deployment of the defect detection methods in manufacturing facilities. The proposed solution aims to provide a general framework for deploying visual defect detection algorithms base on deep neural networks and their high computational requirements. Additionally, the paper addresses challenges related the whole process of automated quality control, which can be performed with varying number of camera devices and it mostly requires interaction with other factory services or workers themselves. We believe the presented framework can contribute to more widespread use of deep learning-based defect detection systems, which may provide valuable feedback for further research and development.
|
|
Implementation of a deep learning model for segmentation of multiple myeloma in CT data
Gálík, Pavel ; Nohel, Michal
This paper deals with the implementation of a deep learning model for spinal tumor segmentation of multiple myeloma patients in CT data. Deep learning is becoming an important part of developing computer-aided detection and diagnosis systems. In this study, a database of 25 patients who were imaged on spectral CT and for whom different parametric images (conventional CT, virtual monoenergetic images, calcium suppression images) were reconstructed, was used. Three convolutional neural network models based on the nnU-Net framework for lytic lesion segmentation were trained on the selected data. The results were evaluated on a test database and the trained models were compared.
|
|
Deep prior audio compression
Švento, Michal ; Balušík, Peter
Audio compression is still an up-to-date topic because the demand for big data streams is rapidly increasing. Deep learning has brought up new algorithms that decrease bitrates with good perception quality. The novel approach in generative artificial intelligence is to produce new data from prior stored in network parameters, called a deep prior. The deep audio prior framework shows its success in various tasks such as inpainting, declipping, and bandwidth extension, but it has not been tested for compression. In this paper, we test this method with a prebuilt network for inpainting. Our idea of compression is based on reducing the number of time-frequency coefficients in the spectrogram while allowing the reconstruction of the original signal with high quality.
|
|
Graph Neural Networks in Epilepsy Surgery
Hrtonová, Valentina ; Filipenská, Marina ; Klimeš, Petr
Epilepsy surgery presents a viable treatment option for patients with drug-resistant epilepsy, necessitating precise localization of the epileptogenic zone (EZ) for optimal outcomes. As the limitations of currently used localization methods lead to a seizure-free postsurgical outcome only in about 60% of cases, this study introduces a novel approach to EZ localization by leveraging Graph Neural Networks (GNNs) for the analysis of interictal stereoelectroencephalography (SEEG) data. A GraphSAGE-based model for identifying resected seizure-onset zone (SOZ) electrode contacts was applied to a clinical dataset comprising 17 patients from two institutions. This study uniquely focuses on the use of interictal SEEG recordings, aiming to streamline the presurgical monitoring process and minimize risks and costs associated with prolonged SEEG monitoring. Through this innovative approach, the GNN model demonstrated promising results, achieving an Area Under the Receiver Operating Characteristic (AUROC) score of 0.830 and an Area Under the Precision-Recall Curve (AUPRC) of 0.432. These outcomes along with the potential of GNNs in leveraging the patient-specific electrode placement highlight their potential in enhancing the accuracy of EZ localization in drug-resistant epilepsy patients.
|
|
Degree of protein structure disorder in prokaryotic and eukaryotic organisms
Nováková, Zuzana ; Vondrášek, Jiří (vedoucí práce) ; Feidakis, Christos (oponent)
Paradigma vztahu mezi strukturou a funkcí proteinů prošlo v posledních třiceti letech revolucí s objevem vnitřně neuspořádaných proteinů (IDP) a regionů (IDR). Tyto proteiny se ukázaly být klíčovými pro řadu buněčných procesů, včetně signalizace, interakcí mezi proteiny a buněčné regulace. Ačkoliv význam IDP/IDR pro funkci živých organismů je nesporný, jejich strukturní analýza představuje významnou výzvu. I přes pokroky v NMR spektroskopii a v algoritmech hlubokého učení pro predikci struktur proteinů zůstávají IDP/IDR stále relativně neznámou oblastí, se značnými mezerami ve znalostech o jejich chování a funkci v živých systémech. Vnitřně neuspořádané proteiny (IDP) se vyskytují ve všech živých organismech, ale jejich hojnost ukazuje na korelaci mezi složitostí organismů a stupněm neuspořádanosti proteinů. Prokaryotické organismy vykazují mnohem nižší výskyt IDPs než eukaryotické. Zvláště významný stupeň neuspořádanosti je pozorován u jednobuněčných parazitických protistů, což naznačuje, že IDP mají zásadní význam v patogenezi a průběhu nemocí jako je malárie a toxoplazmóza. U lidí jsou dysfunkce IDP spojeny s mnoha onemocněními, včetně neurodegenerativních chorob, jako je Parkinsonova a Alzheimerova nemoc, jako různé typy rakoviny. Porozumění těmto proteinům by mohlo významně ovlivnit vývoj...
|
|
Methods for Realtime Voice Deepfakes Creation
Alakaev, Kambulat ; Pleško, Filip (oponent) ; Malinka, Kamil (vedoucí práce)
This thesis explores the possibility of achieving real-time voice deepfake generation using open-source tools. Through experiments, it was discovered that the generation rate of voice deepfakes is affected by the computing power of the devices running the speech creation tools. A deep learning model was identified to be capable of generating speech in near real time. However, limitations in the tool containing this model prevented continuous input data for real-time generation. To address this, a program was developed to overcome these limitations. The quality of the generated deepfakes was evaluated using both voice deepfake detection models and human online surveys. The results revealed that while the model could deceive detection models, it was not successful in fooling humans. This research highlights the accessibility of open-source voice synthesis tools and the potential for their misuse by individuals for fraudulent purposes.
|
|
Detecting Presentation Attacks Using Face Spoofing
Homola, Tomáš ; Orság, Filip (oponent) ; Goldmann, Tomáš (vedoucí práce)
Face detection is one of the most important and widespread methods of verifying a person's identity. However, this method also raises concerns about privacy and security. It is important to be aware of the dangers it brings and constantly develop the necessary means to protect against them. This thesis aims to explain the issue of face spoofing, the threat that arises from a successful attacker's attempt at spoofing, and the detection of these spoofs using algorithms.
|
|
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.
|
|
Systém pro rozpoznávání dezinformací v prostředí webu
Večerka, Lukáš ; Žádník, Martin (oponent) ; Strnadel, Josef (vedoucí práce)
Tato práce se zabyvá návrhem, realizací a ověřením systému pro automatické rozpoznávání dezinformací v prostředí webu. Představuje problematiku šíření dezinformací v online prostředí a jeho dopad na společnost. Zaměřuje se na trénování několika Českych transformers jazykovych modelů pro rozpoznání dezinformací a dále na automatickou extrakci obsahu článků z českych internetovych novin a jejich analyzu využitím klasifikace textu a zpracování přirozeného jazyka pomocí metod hlubokého učení. Vysledky těchto analyz jsou pak prezentovány na webovém uživatelském rozhraní s cílem poskytnout platformu pro ověření článků, autorů a zdrojů. Rozhraní by mohlo byt použito k anotaci dat experty pro průběžné vylepšování jazykovych modelů.
|
| |