|
|
Using machine learning for quality control in industrial applications
Gaško, Viktor ; Dobrovský, Ladislav (referee) ; Parák, Roman (advisor)
Goal of this bachelor´s thesis is to get acquainted with issue of quality control in industrial applications with focus on deep learning. For this and similar issues was created several libraries which have a purpose of simplifying these issues. Main task is to create program for quality control with help of programming language Python and framework Tensorflow. This program will be comprised of three neural network, from which one will identify the approximate position of the part, second its color, and third will check the correctness of its production.
|
|
|
Shared Experience in Reinforcement Learning
Mojžíš, Radek ; Šůstek, Martin (referee) ; Hradiš, Michal (advisor)
The aim of this thesis is to use methods of transfer learning for training neural network on a reinforcement learning tasks. As test environment, I am using old 2D console games, such as space invaders or phoenix. I am testing the impact of re-purposing already trained models for different environments. Next I use methods for domain feature transfer. Lastly i focus on the topic of multi-task learning. From the results we can gain insight into possibilities of using transfer learning for reinforcement learning algorithms.
|
|
|
Fine-Grained Recognition and Re-Identification of Vehicles Using Advanced Feature Extraction
Doseděl, Ondřej ; Hradiš, Michal (referee) ; Špaňhel, Jakub (advisor)
Práce se zabývala analýzou a následným vylepšením metod užívaných k rozpoznávání typů vozidel a jejich re-identifikace. Navržená metoda může být využita jak pro rozeznání, tak pro re-identifikaci. Byla založena na používání tzv. 3D bounding boxes. Pomocí těchto boxů docházelo k detekci vozidla na obraze. Vozidlo bylo následně normováno rozbalením do dvojrozměrné interpretace. Tato metoda byla vylepšena určením směru vozidla a rozlišováním mezi čelní a zadní stranou vozidla během rozbalení třírozměrného modelu. Představená metoda vylepšuje stávající metodu pro rozpoznávání a snižuje její chybovost až o 13 % pro jeden vzorek a o 17% pro přesnost trati. Pro re-identifikaci nedošlo k zlepšení při použití LFTD agregovaní.
|
|
|
Deep Learning for Object Detection
Paníček, Andrej ; Herout, Adam (referee) ; Teuer, Lukáš (advisor)
This work deals with the object detection using deep neural networks. As part of the solution, I modified, implemented and trained the well-known model of cascade neural networks MTCNN so that it could perform the detection of traffic signs. The training data was generated from GTSRB and GTSDB data sets. MTCNN showed solid performance on the evaluation data, where the detection accuracy reached 97.8 %.
|
|
|
Detection of Traffic Signs in Image and Video
Kočica, Filip ; Hradiš, Michal (referee) ; Herout, Adam (advisor)
This thesis deals with the traffic sign detection problematics using modern techniques in image processing. Special architecture of deep convolutional neural network YOLO, i.e. You Only Look Once, which performs both detection and classification in one step, has been used. This architecture allows object detector to work on very high speeds. This thesis also deals with comparison of models trained on real and synthetic datasets. The best model trained on real dataset has reached 63.4% mAP success rate and 82.3% mAP when trained on synthetic dataset. Evaluation of one image takes about ~40.4ms on average graphics processing unit and ~3.9ms on higher than average graphics processing unit. The benefit of this thesis is that under certain conditions neural network model trained on synthetic data can achieve same or even better results than model trained on real data. This may simplify process of object detector development since it is not necessary to annotate large number of images.
|
|
|
Detection Of Anatomical Structures In Ct Data
Using Convolutional Neural Networks
Kozlová, Dominika
This paper deals with a detection of anatomical structures in medical images using convolutional neural networks (CNN). The designed algorithm contains 2 methods for region proposals and CNN for their classification into categories. Output of the CNN is then postprocessed to obtain the detection result. Categories for detection are head, spine, heart, left and right lung, aorta, liver, left and right kidney, spleen and background. For training and validation of the network were created 2 sets of CT data with annotated areas of selected structures.
|
|
|
Fine-Grained Vehicle Recognition from Traffic Surveillance Camera
Mencner, Pavel ; Špaňhel, Jakub (referee) ; Sochor, Jakub (advisor)
The aim of this thesis is image based detection of vehicles from traffic surveillance camera and fine-grained vehicle type recognition (manufacturer and model). In the thesis the Unpack normalization method is implemented which transforms the vehicle image into its apparent flat representation in order to increase the classifier's success rate. The Unpack method make use of 3D bounding box of the vehicle. This bounding box is constructed during test period using the information of vehicle contour and direction toward vanishing points. The thesis involve accuracy comparison between direct and Unpack classification methods. The proposed solution is based on several related parts that benefit from convolutional neural networks. These parts are: vehicle detection from image data, estimation of the directions towards vanishing points solved as classification task, vehicle contour detection using convolutional Encoder-Decoder network and fine-grained vehicle type classification. Using Unpack based classification the 2% accuracy improvement against direct classification has been achieved, resulting in 86% overall success rate. The outcome of this thesis is fine-grained vehicle classification system that works with traffic surveillance video without any viewpoint limitations.
|
|
|
Detection of Weapons in 2D Image
Demčák, Ján ; Spurný, Martin (referee) ; Drahanský, Martin (advisor)
This bachelor thesis deals with detection of weapons in 2D image. In the theoretical part of the thesis the term weapon was defined and the possibilities of detection of weapons in image with using classic methods and deep neural networks were mentioned there. The key steps of image processing, objects classification and detection were described. The overview of frameworks, libraries was presented. To implement the pratical part of the thesis, 3 models were chosen. The first classic model with using HOG transformation. The second CNN model with priority target detection accuracy and with two different neural network architectures as classifiers. The third model with YOLO network architecture had as priority target real-time detection. The essential part of each model was choice, or more precisely creating suitable dataset. What followed was the construction and implementation of models and the evaluation of obtained data.
|
|
|
Navigation Using Deep Convolutional Networks
Skácel, Dalibor ; Veľas, Martin (referee) ; Hradiš, Michal (advisor)
In this thesis I deal with the problem of navigation and autonomous driving using convolutional neural networks. I focus on the main approaches utilizing sensory inputs described in literature and the theory of neural networks, imitation and reinforcement learning. I also discuss the tools and methods applicable to driving systems. I created two deep learning models for autonomous driving in simulated environment. These models use the Dataset Aggregation and Deep Deterministic Policy Gradient algorithms. I tested the created models in the TORCS car racing simulator and compared the result with available sources.
|
|
|
Neural Network Implementation without Multiplication
Slouka, Lukáš ; Baskar, Murali Karthick (referee) ; Szőke, Igor (advisor)
The subject of this thesis is neural network acceleration with the goal of reducing the number of floating point multiplications. The theoretical part of the thesis surveys current trends and methods used in the field of neural network acceleration. However, the focus is on the binarization techniques which allow replacing multiplications with logical operators. The theoretical base is put into practice in two ways. First is the GPU implementation of crucial binary operators in the Tensorflow framework with a performance benchmark. Second is an application of these operators in simple image classifier. Results are certainly encouraging. Implemented operators achieve speed-up by a factor of 2.5 when compared to highly optimized cuBLAS operators. The last chapter compares accuracies achieved by binarized models and their full-precision counterparts on various architectures.
|
guest ::
