Synthesis of normal map textures for physically based rendering materials with artificial neural networks

Procedural Content Generation (PCG) is a field that has been studied on for decades. PCG may make it possible to lower costs of digital products. Many digital contents include commonly-used or very similar assets. Finding a proper function for PCG would decrease time and money required for the production. On the other hand, preparing useful functions to generate digital assets is difficult. This kind of functions can be approximated using machine learning methods. Given a sufficient amount of data, a generative artificial intelligence model can be trained to produce a specific type of asset. Thank to the improvements on Graphics Processing Units (GPU), it is easier nowadays to build artificial intelligence models. Especially, Artificial Neural Networks (ANN) have became popular and been worked on actively for years. Training models that can classify or produce samples is of the capabilities of ANN's. Autoencoders and Generative Adverserial Networks (GAN) are examples of generative models. Using generative models, it is also possible to transfer styles or approximate other complex functions. Physically Based Rendering (PBR) method is a way to render realistic 3D models and scenes in realtime. For good results, various type of textures are needed. Bump maps are a type of texure for 3D materials. They are used to simulate surface normals and heights. Normal maps are bump maps that store normal vectors of surface. They can be produced using photometric stereo techniques or projecting a high-poly model onto its low-poly version. These methods are time consuming and expensive. There are many studies that focus on combining ANN and PCG. Using both, it would be possible to produce material textures in reasonable quality while reducing the costs. In this thesis, several generative artificial neural network models were trained. Their results were compared to eachother by the network complexity. It was determined that increasing layer and filter numbers doesn't always higher the output quality. The complex models caused vanishing gradient problem. Vanishing gradient problem, at some point, stops the model from getting more improved and makes it ineffective to continue the training phase. It was understood that the models with the architecture of Residual Neural Networks (ResNet) can handle this problem. Final results also showed that data augmentation plays a big role on the training phase.