Neural networks are trending topic at the moment. More and more problems are being solved with this type of approach. At the beginning, these implementations could solve problems only with smaller datasets. The effective use of video cards is one of the key elements for the success of neural nets. We can train more data with less time, but the most important improvement is that the computational power opened the door for solving various new problems. One of these problems is image processing.
The thesis deals with deep neural network-based image generating methods. For image classification there are a few very good solutions, but this problem is the straight opposite. We generate images from random data or labels. There are many ways to achieve this goal. There are three model descriptions in the thesis, one of which is implemented. The model has been implemented in two forms. One of the is for weaker video cards, the second contains more trainable parameters and can only be taught with video cards that have more memory than 2GB. I measured the image quality and the training time. The smaller network was trained with both video cards.