Dezvoltarea sistemelor inteligente de interfațare vizuală om-mașină / The development of intelligent systems for visual human-machine interfaces

The substantial increase in performance of Machine Learning systems which was noticed in the last decades can be explained by a series of independent factors: the development of the various used algorithms, the increase in both quality and quantity of data, etc. The main target of the research which was conducted in preparation for this paper was the improvement of Convolutional Neural Networks. Although the performance of these networks has made them the most common choice for applications involving visual information, the training process is a costly one from the perspective of computational resources and they require larger quantities of data to ensure good results, when compared to other algorithms. The methods used to gain the extra performance were focused on the augmentation of the datasets or the enhancement of the learning algorithms, by employing new loss functions to be used alongside the typical ones. From the perspective of the target applications, two major research directions are noted. The first one involved the analysis of paintings, the main task being genre recognition. One of the more difficult aspects of this domain is the varying degree of abstraction which can be noticed when analyzing visual art. The second research theme was the analysis of facial expressions. In this case, multiple methods of representation were considered, some of the solutions employing the specificities of the theme. Besides the regular approaches specific to supervised learning, some of the investigated enhancement belong to the area of semi-supervised learning, managing to make use of data which is not annotated, but more easily accessible.