Department of Mathematical Informatics

v_ter@ukr.net
(044)259-04-27

Zinchenko A. B. Description images using recurrent neural networks.

Master’s thesis is devoted to the problem of describing images. The main aim of the thesis – to show that the technique of deep learning can be used to solve a wide class of problems, including the problem description and images. In the thesis work the architecture of neural networks. The final network consisted of convolutional and recurrent GRU layers of interconnected cells. To train network used the method of transfer options. At the end of the work was the analysis results. The final network has shown impressive results. This is to prove that neural networks are able to solve very complex and unusual problems. Also, these networks have a single set of parameters for optimization that significantly speeds development and increases the quality of such models.