Nowdays with the help of data science computers are able to solve an increasing number of intelligent tasks, such as object recognition on images or different kinds of image classification problems. The goal of this thesis was to create a system which can learn the classification of images using labels containing semantic information, and classify an unknown set of images based on that model. The system also had to be suited with the ability to separate images of completely unknown classes, i.e. to recognize classes that did not appear during training.
In this paper, I summarize the steps of a possible solution to image representation that I use to create high-level descriptors for image classification algorithms from descriptors containing low-level semantic information of images. I introduce a number of different classification algorithms frequently used in the world of data mining that are also well suited for image classification problems, while I also introduce techniques used in Open Set Recognition, which deals with unknown classes during training. At the end of this paper, I use the the implemented system to compare the previously introduced classifiers in a number of different terms.