Detekce objektů s využitím neuronových sítí

Abstract

This thesis deals with and solves the problem of image detection using neural networks. The thesis describes a broader issue from image recognition, through the principles of neural and convolutional neural networks. Selected algorithms and architectures of convolutional neural networks were investigated and subsequently described. For the practical part, a popular and special architecture of a deep convolutional neural network, called YOLO (You Only Look Once), was used, which is different from the other architectures in that the detection and classification takes place in one step. Furthermore, the work describes some possibilities of network training on own hardware or on external data servers. In the next part of the work, the YOLO convolution network was trained on various data sets for the detection of one or more objects. The work compares the success of trained models for different network parameters.