|
|
Evolutionary Design of Image Classifier
Koči, Martin ; Bidlo, Michal (referee) ; Drahošová, Michaela (advisor)
This thesis deals with evolutionary design of image classifier with help of genetic programming, specifically with cartesian genetic programming. Thesis discribes teoretical basics of machine learing, evolutionary algorithms and genetic programming. Part of this thesis is described design of the program and its implementation. Futhermore, experiments are performed on two solved tasks for the classification of handwritten digits and the classification of cube drawings, which can be used to determine the rate of dementia in Parkinson's disease. The best designed solution for digits is with AUC of 0.95 and for cubes 0.86. Designed solutions are compared by other methods, namely convolutional neural networks (CNN) and the support vector machines (SVM). The resulting AUC for the classification of digits for both CNN and SVM is 0.99, for cubes CNN has a final AUC 0.81 and SVM 0.69. The cubes are then compared with existing solution, which resulted in AUC 0.70, so that the results of the experiments show an improvement in the method used in this thesis.
|
|
|
Application of Neural Accelerators on Rapsberry PI
Barna, Kristian ; Sekanina, Lukáš (referee) ; Vašíček, Zdeněk (advisor)
The presented bachelor thesis deals with the statistical evaluation of performance for hardward accelerator of deep neural networks. Describes convolutional neural networks along with mathematical calculations. Explains their acceleration and conversion to a format suitable for the Intel Movidius NCS accelerator. 8 hardware platforms and 22 neural network difficulties were compared experimentally. Up to 105-fold improvement was demonstrated in isolated inference of the MobileNetV2 network for the Raspber Pi platform using an accelerator. Performance between the tested platforms was also evaluated from an energy point of view. The application of facial identity demonstrated the conditions during real use. Possible limits of CNN acceleration on power-limited devices (Raspberry Pi) have been uncovered, especially due to improper selection of input image resolution. All measurements were evaluated by statistical procedures.
|
|
|
Evaluation of targets in shooting range based on image data
Sujová, Sára ; Šťastný, Jiří (referee) ; Škrabánek, Pavel (advisor)
The thesis describes the design and implementation of a computer vision system for evaluating targets on the shooting range using image data. The program respects the restrictions based on safety measures established by the the shooting range manager and uses an uniform system of lighting and camera placement. The work consists of several parts. The first part is the creation of the dataset and its annotation. The second part is the creation of the program. The program includes a photo of the target, which is suitably edited and divided into sub-areas in the pre-processing phase. These sub-regions are then iteratively processed by the U-NET network, which produces segmentation maps that are subsequently combined into the resulting map. The positions of the detected shots are obtained from this map. In the last part of the program, a point evaluation of the shooting session is obtained.
|
|
|
Object Detection in the Laser Scans Using Convolutional Neural Networks
Marko, Peter ; Beran, Vítězslav (referee) ; Veľas, Martin (advisor)
This thesis is aimed at detection of lines of horizontal road markings from a point cloud, which was obtained using mobile laser mapping. The system works interactively in cooperation with user, which marks the beginning of the traffic line. The program gradually detects the remaining parts of the traffic line and creates its vector representation. Initially, a point cloud is projected into a horizontal plane, crating a 2D image that is segmented by a U-Net convolutional neural network. Segmentation marks one traffic line. Segmentation is converted to a polyline, which can be used in a geo-information system. During testing, the U-Net achieved a segmentation accuracy of 98.8\%, a specificity of 99.5\% and a sensitivity of 72.9\%. The estimated polyline reached an average deviation of 1.8cm.
|
|
|
Vehicle License Plate Detection and Recognition Software
Masaryk, Adam ; Hradiš, Michal (referee) ; Špaňhel, Jakub (advisor)
The aim of this bachelor thesis is to design and develop software that can detect and recognize license plates from images. The software is divided into 3 parts - license plates detection, detector output processing and license plates characters recognition. We decided to implement detection and recognition using modern methods using convolutional neural networks.
|
|
|
Comic Images Super-Resolution Using Deep Learning
Zdravecký, Peter ; Juránek, Roman (referee) ; Španěl, Michal (advisor)
Táto práca demonštruje metódu super rozlíšenia na zlepšenie kvality komiksových obrázkov pomocou hlbokého učenia. Náročnou časťou tejto úlohy bolo súčasne zachovať kvalitu textových a kreslených častí, bez výraznej deformácie ktorejkoľvek časti z nich. Na dosiahnutie uspokojivých výsledkov boli skúmané dve hlboké neurónové siete. Sieť U-Net a modifikácia s názvom Robustný U-Net (RUNet). Zvolené stratové funkcie na trénovanie týchto sietí boli stredná kvadratická chyba a perceptuálna strata. Práca obsahuje experimenty na týchto sieťach v kombinácii s každou stratovou funkciou. Ďalšie experimenty sa zamerali na vplyv počtu použitých blokov zo stratovej siete VGG16 na funkciu perceptuálnej straty. Experimenty ukázali, že sieť RUNet využívajúca perceptuálnu stratu s tromi extrahovanými blokmi dosiahla najlepšie výsledky.
|
|
|
Pedestrian Attribute Analysis
Studená, Zuzana ; Špaňhel, Jakub (referee) ; Hradiš, Michal (advisor)
This work deals with obtaining pedestrian information, which are captured by static, external cameras located in public, outdoor or indoor spaces. The aim is to obtain as much information as possible. Information such as gender, age and type of clothing, accessories, fashion style, or overall personality are obtained using using convolutional neural networks. One part of the work consists of creating a new dataset that captures pedestrians and includes information about the person's sex, age, and fashion style. Another part of the thesis is the design and implementation of convolutional neural networks, which classify the mentioned pedestrian characteristics. Neural networks evaluate pedestrian input images in PETA, FashionStyle14 and BUT Pedestrian Attributes datasets. Experiments performed over the PETA and FashionStyle datasets compare my results to various convolutional neural networks described in publications. Further experiments are shown on created BUT data set of pedestrian attributes.
|
|
|
Volumetric Segmentation of Dental CT Data
Berezný, Matej ; Kodym, Oldřich (referee) ; Čadík, Martin (advisor)
The main goal of this work was to use neural networks for volumetric segmentation of dental CBCT data. As a byproducts, both new dataset including sparse and dense annotations and automatic preprocessing pipeline were produced. Additionally, the possibility of applying transfer learning and multi-phase training in order to improve segmentation results was tested. From the various tests that were carried out, conclusion can be drawn that both multi-phase training and transfer learning showed substantial improvement in dice score for both sparse and dense annotations compared to the baseline method.
|
|
|
Evolutionary Algorithms in Convolutional Neural Network Design
Badáň, Filip ; Vašíček, Zdeněk (referee) ; Sekanina, Lukáš (advisor)
This work focuses on automatization of neural network design via the so-called neuroevolution, which employs evolutionary algorithms to construct artificial neural networks or optimise their parameters. The goal of the project is to design and implement an evolutionary algorithm which can be used in the process of designing and optimizing topologies of convolutional neural networks. The effectiveness of the proposed framework was experimentally evaluated on tasks of image classification on datasets MNIST and CIFAR10 and compared with relevant solutions. The results showed that neuroevolution has a potential to successfully find accurate and effective convolutional neural network architectures.
|
|
|
Interpretability of Neural Networks in Speech Processing
Sarvaš, Marek ; Mošner, Ladislav (referee) ; Žmolíková, Kateřina (advisor)
S rastúcou popularitou hlbokých neurónových sietí, nedostatok transparentnosti spôsobenejich funkciou čiernej skrinky, zvyšuje dopyt po ich interpretácii. Cieľom tejto práce je získať nový pohľad na hlboké neurónové siete v úlohách spracovania reči. Konkrétne klasifikácia pohlavia z AudioMNIST datasetu a klasifikácia rečníka z filter bánk VoxCeleb datasetu s použitím konvolučnej a reziduálnej neurónovej siete. Na interpretáciu týchto neurónových sietí bola použitá metóda propagácie relevancií cez vrstvy. Táto metóda vytvorí tepelnú mapu, ktorá vyznačí príznaky, ktoré prispeli ku správnej klasifikácii pozitívne a ktoré negatívne. Ako výsledky interpretácie ukazujú, klasifikácie boli založené najmä na nižších frekvenciách v reči a čase. V prípade klasifikácie pohlavia sa mi podarilo nájsť vysokú závislosť modelu na veľmi malom počte príznakov. Pomocou získaných informácií som vytvoril rozšírený trénovací set, ktorý zvýšil robustnosť modelu.
|
guest ::
