During the past decade the huge advancements in terms of computational power and resources implied the exponential development of the different hardware demanding machine learning techniques. One of the most popular area in this field are the different, biologically inspired microscopic image processing methods. Here two important steps are finding the proper focus plane for a sharp image and additionally segmenting or detecting the most important areas (objects) on the images. This is the

more »

... tion also with the Digital holographic microscopy (DHM), which is a proper technique for the imaging of three dimensional objects, because — compared to conventional microscopes — it has better depth-of-field through allowing the reconstruction of a volume with multiple in-focus planes. In this thesis work I propose deep learning based solutions to firstly find the focus planes of the objects on the DHM recordings, and then segment their locations with their types as well. For these, I implemented different kinds of convolutional neural networks using the Python programming language and the PyTorch framework. Regarding the focus distance estimation, after the architecture of the network has been designed and implemented, 3 different setup was inspected for finding an optimal solution: The target distances (direction - positive and negative distances included, absolute - absolute value of the distances), the activation function in the last layer (hyperbolic tangent or sigmoid) and the inclusion or exclusion of the phase images into the input of the net. Concerning the image segmentation, the task here was more complex. Here the available labelled images for the ground truth were not accurate enough, thus I firstly developed an image processing tool, which had two main functionalities: image augmentation through transformations and semi-automatic image labelling. Then, as the next step a simple object segmentor net (SOSN) was impl [...]