Collection of big datasets is extremely challenging for many fields (e.g healthcare and biology) and goals have to be achieved with limited amount of data. Data Augmentation (DA) techniques can help us mitigate this issue. The recent rise of Deep Generative Models has made them very attractive models to perform DA. In this Master Thesis we explore the field of Deep Generative Modelling of Microscopy Image Data that depict the Arbuscular Mycorrhiza Fungi (AMF). We develop a Variational Autoencoder (VAE) in order to generate very similar synthetic Microscopy Image Data which can later be used to increment the size of AMF datasets and allow them to be utilized further for classification or segmentation purposes. We provide a full generation pipeline, where the original dataset is being preprocessed, split into smaller parts, passed to the model for training and used afterwards for generating new image data. The VAE has a new architecture, developed from scratch and inspired by Residual Network (ResNet). Additionally, a novel training approach for the ResNet based VAE has been developed and explored which is inspired by Progressive Generative Adversarial Network (ProGAN). The results are encouraging and motivate future research in the field of Deep Generative Modelling.     «

Collection of big datasets is extremely challenging for many fields (e.g healthcare and biology) and goals have to be achieved with limited amount of data. Data Augmentation (DA) techniques can help us mitigate this issue. The recent rise of Deep Generative Models has made them very attractive models to perform DA. In this Master Thesis we explore the field of Deep Generative Modelling of Microscopy Image Data that depict the Arbuscular Mycorrhiza Fungi (AMF). We develop a Variational Autoencode...     »