|
|
Využití akcelerátoru neuronových sítí na Raspberry PI
Barna, Kristian ; Sekanina, Lukáš (oponent) ; Vašíček, Zdeněk (vedoucí práce)
Predkladaná bakalárska práca sa zoberá štatistickým vyhodnotením výkonnosti hardwardového akcelerátora hĺbkových neurónových sieti. Opisuje Konvolučné neurónové siete spolu s matematickými výpočtami. Vysvetľuje ich akceleráciu a prevod do formátu vhodného pre akcelerátor Intel Movidius NCS. Experimentálne sa porovnalo 8 hardvérových platforiem a 22 náročností neurónovej sieti. Bolo demonštrované až 105 násobné zlepšenie pre izolovanú inferenciu MobileNetV2 siete na platformu Raspbery Pi za pomoci akcelrátora. Výkon medzi testovanými platformami sa vyhodnocoval aj z energetického hľadiska. Aplikáciou na rozpoznanie identity tváre sa demonštrovali podmienky reálneho použitia. Odkryli sa možné limity akcelerácie CNN na zariadeniach s obmedzeným výkonom (Raspberry Pi), najmä v súvislosti s nevhodným výberom vstupného rozlíšenia obrazu. Všetky merania boli vyhodnocované štatistickými postupmi.
|
|
|
Hluboké neuronové sítě pro analýzu medicínských obrazových dat
Szöllösi, Albert ; Kodym, Oldřich (oponent) ; Španěl, Michal (vedoucí práce)
Táto práca ponúka možné riešenie automatickej lokalizácie anatomických landmarkov v 3D dentálnych skenoch. Tieto skeny slúžia k uľahčeniu modelovania zubných koruniek a ortodontických aparátov pomocou špecializovaného softvéru. Predtým však musia byť anotované, aby softvér vedel určiť, kde sa jednotlivé zuby nachádzajú. Anotácia prebieha ručne, čo síce zaručuje presnosť, ale zaberá veľa času. Výsledok tejto práce by mohol spomínaný postup výrazne zjednodušiť aplikovaním hlbokého učenia. Lokalizácia landmarkov bola riešená pomocou konvolučnej neurónovej siete.
|
|
|
Deep Learning Algorithms on Embedded Devices
Hadzima, Jaroslav ; Boštík, Ondřej (oponent) ; Horák, Karel (vedoucí práce)
This paper describes currently widely used Deep Learning architectures and methods for object detection and classification in video, with intention of using them on embedded systems. We will cover steps and reasoning when choosing the most appropriate embedded hardware for our application. Our test application consists of vehicle detection and free parking space detection using Deep learning methods, all wrapped under name Smart car park. This application provides monitoring of vehicle presence in car park and if they occupy parking spot or not. All this is expected to be done using embedded device. Later, there will be covered configuration steps for our embedded device with emphasis on hardware optimization for speed. We will provide comparison of available inference models, which will be rated mostly in categories like speed or F1 score, which have the biggest impact in our application. The best candidate will be selected and used for testing of our application.
|
|
|
Detekce objektů v laserových skenech pomocí konvolučních neuronových sítí
Marko, Peter ; Beran, Vítězslav (oponent) ; Veľas, Martin (vedoucí práce)
Práca sa zaoberá detekciou čiary vodorovného dopravného značenia z mračna bodov, ktoré bolo získané laserovým mobilným mapovaním. Systém pracuje interaktívne v spolupráci s užívateľom, ktorý vyznačí počiatok čiary dopravného značenia. Program postupne deteguje zvyšné časti dopravného značenia a vytvorí ich vektorovú reprezentáciu. Na začiatku je mračno bodov premietnuté do vodorovnej roviny a výsledkom je 2D obrázok, ktorý je segmentovaný konvolučnou neurónovou sieťou U-Net. Segmentácia označuje jednu dopravnú čiaru. Segmentácia je prevedená na lomenú čiaru, ktorú je možné použiť v geo-informačnom systéme. Sieť U-Net pri testovaní dosiahla presnosť segmentácie 98,8\%, špecificitu 99,5\% a senzitivitu 72,9\%. Odhadnutá lomená čiara dosiahla priemernú odchýlku 1,8cm.
|
|
|
Segmentation of multiple sclerosis lesions using deep neural networks
Sasko, Dominik ; Myška, Vojtěch (oponent) ; Kolařík, Martin (vedoucí práce)
This master thesis focused on automatic segmentation of Multiple Sclerosis (MS) lesions on MRI images. We tested the latest methods of segmentation using Deep Neural Networks and compared the approaches of weight initialization by transfer learning and self-supervised learning. The automatic segmentation of MS lesions is a very challenging task, primarily due to the high imbalance of the dataset (brain scans usually contain only a small amount of damaged tissue). Another challenge is a manual annotation of these lesions, as two different doctors can mark other parts of the brain as damaged and the Dice Coefficient of these annotations is approximately 0.86, which further underlines the complexity of this task. The possibility of simplifying the annotation process by automatization could improve the lesion load determination and might lead to better diagnostic of each individual patient. Our goal was to propose two techniques that use transfer learning to pre-train weights to later improve the performance of existing segmentation models. The theoretical part describes the division of artificial intelligence, machine learning and deep neural networks and their use in image segmentation. Afterwards, the work provides a description of Multiple Sclerosis, its types, symptoms, diagnosis and treatment. The practical part begins with data preprocessing. Firstly, brain scans were adjusted to the same resolution with the same voxel size. This was needed due to the usage of three different datasets, in which the scans had been created by devices from different manufacturers. One dataset also included the skull, therefore it was necessary to remove it by an FSL tool, leaving only the patient's brain in the scan. The preprocessed data were 3D scans (FLAIR, T1 and T2 modalities), which were cut into individual 2D slices and used as an input for the neural network with encoder-decoder architecture. The whole dataset consisted a total of 6,720 slices with a resolution of 192 x 192 pixels for training (after removing slices where the mask was empty). Loss function was Combo loss (combination of Dice Loss with modified Cross-Entropy). The first technique was to use the pre-trained weights from the ImageNet dataset on encoder in U-Net network, with and without locked encoder weights, respectively, and compare the results with random weight initialization. In this case, we used only the FLAIR modality. Transfer learning has proven to increase the metrics from approximately 0.4 to 0.6. The difference between encoder with and without locked weights was about 0.02. The second proposed technique was to use a self-supervised context encoder with Generative Adversarial Networks (GAN) to pre-train the weights. This network used all three modalities also with the empty slices (23,040 slices in total). The purpose of GAN was to recreate the brain image, which was covered by a checkerboard. Weights learned during this training were later loaded for the encoder to apply to our segmentation problem. The following experiment did not show any improvement, with a DSC value of 0.29 and 0.09, with and without a locked encoder, respectively. Such a decrease in performance might have been caused by the use of weights pre-trained on two distant problems (segmentation and self-supervised context encoder) or by difficulty of the task considering the hugely unbalanced dataset.
|
|
|
Detection and evaluation of distorted frames in retinal image data
Vašíčková, Zuzana ; Chmelík, Jiří (oponent) ; Kolář, Radim (vedoucí práce)
The master's thesis deals with detection and evaluation of distorted frames in retinal image data. The theoretical part contains brief summary of eye anatomy and methods for image quality assessment generally, and also particularly on retinal images. The practical part is carried out in programming language Python. It contains preprocessing of the available retinal images in order to create an appropriate dataset. Further a method for evaluation of three types of blur in distorted retinal images is proposed, specifically Inception-ResNet-v2 model. This method is not feasible and thus another method consisting of two steps is designed - classification of the type of blur and subsequently evaluation of the particular blur level. Filtered Fourier spectrum is used to classify the type of blur and features extracted by ResNet50 serve as the input for regression model. This method is further extended with initial step of detection of blurred frames in retinal sequences.
|
|
|
Horizon Detection in Image
Holková, Natália ; Herout, Adam (oponent) ; Juránek, Roman (vedoucí práce)
This thesis aims to implement a method of detecting the horizon line in images using deep learning to prevent any constraints on input data. A training dataset is created by downloaded images from large metropolitan cities around the world using the Google Street View service. Several popular architectures for convolutional neural networks are chosen, and their performance is evaluated on existing benchmark datasets.
|
|
|
Comic Images Super-Resolution Using Deep Learning
Zdravecký, Peter ; Juránek, Roman (oponent) ; Španěl, Michal (vedoucí práce)
This paper demonstrates a super-resolution method for improving the resolution and quality of comic images by using deep learning. The challenging part of the task was to keep the quality of the text parts and drawings simultaneously, without significant deformation of any part. Two deep neural networks were used to achieve satisfying results. U-Net network and its modification called Robust U-Net. The chosen loss functions to train these networks were the Mean Squared Error and Perceptual loss. The work contains experiments on U-Net and modified RUNet networks with a combination of each loss function. Additional experiments looked at how the number of used blocks from the VGG16 loss network affects the Perceptual loss function. Experiments have shown that a Robust U-Net network using a Perceptual loss with three extracted blocks got the best results.
|
|
|
Segmentace zubních objemových dat
Berezný, Matej ; Kodym, Oldřich (oponent) ; Čadík, Martin (vedoucí práce)
Hlavným cieľom tejto práce bola segmentácia objemových CT dát za použitia neurónových sietí. Ako vedľajší produkt bol vytvorený nový dataset spolu s silnými aj slabými anotáciami a nástroj pre automatický preprocessing dát. Takisto bola overená možnosť využitia transfer learningu a viacfázového trénovania. Z mnohých vykonaných testov možno vyvodiť záver, že aj tranfer learning aj viacfázové trénovanie mali pozitívny vplyv na vývoj dice skóre v porovnaní so základnou použitou metódou či už pri silných, alebo slabých anotáciách.
|
|
|
Deep Neural Networks Used for Customer Support Cases Analysis
Marušic, Marek ; Ryšavý, Ondřej (oponent) ; Pluskal, Jan (vedoucí práce)
Artificial intelligence is remarkably popular these days. It can be used to resolve various highly complex tasks in fields such as image processing, sound processing, natural language processing, etc. Red Hat has an extensive database of resolved support cases. Therefore an idea was proposed to use these data for data mining and information retrieval in order to ease a resolution process of the support cases. In this work, various deep neural network models were created for prediction of features which could help during the resolution process. Techniques and models used in this work are described as well as their performance in the specific tasks. Comparison of individual models is outlined as well.
|
host ::
RSS.