Generation of face images using deep convolutionary advertising generative networks

In this study, the increase of knowledge from primitive man to the present day, its intergenerational transfer and the transformation of this process into big data are examined. At the same time, topics such as the historical development of artificial intelligence, the structure of artificial neural networks, deep learning, deep learning layers and deep learning models are discussed in detail. In addition, detailed information is presented about adversarial generative networks, which is one of the image generation models and a sub-branch of deep learning. The historical process of contentious productive networks, academic studies, their development and contributions to academic literature have been meticulously researched. In particular, the DCGAN model, which is one of the best models in the field of image generation of adversarial generative networks, has been examined in detail. An intensive study has been carried out on the structure of the generative network that forms the DCGAN model, its mathematical methods and the functions used in the experiments, as well as the loss function values. In addition, the structure, functions, mathematical methods and functions of the discriminator network, which is the second component of adversarial generator networks, and the functions used in the experiments are discussed in detail. The libraries, GPU, processor, language processing editors and Google Colab environment used in the experimental environment were carefully examined. The most commonly used datasets to generate synthetic images have been carefully selected. Four data sets of two different types were used in this study. First, Cartoonset10k and Anime Face datasets containing vector-based images were carefully selected and analyzed in detail. Secondly, Animal Face and CelebaFace datasets containing pixel-based images were examined in detail. In the experiments carried out on each data set, the 8x8 images obtained as a result of the training processes repeated at regular intervals were recorded step by step. These best images were evaluated in detail by comparing them with randomly selected images in the data sets. Loss-gain values of the generating network and the discriminator network according to Nash equilibrium were obtained using mathematical and graphical methods.