Neural network based tools enjoy a renaissance in the field of image processing these days. The renewed interest in this field of research has been supported by widely accessible hardware with appropriate computational capacity. Although tools of the neural paradigm, surpassing solutions of the classic methods, have proved the versatility of their applicability in several fields, intrinsic difficulties of the method represent a challenge. Consequently, in some cases the choice between classic image processing or neural paradigm based approaches is not undisputed even though proliferating publications on neural network based approaches that are successful beyond past solutions threaten to flood the literature.
Accordingly, my thesis focuses on the review and comparison of classic methods and neural net based solutions highlighting the theoretical background of the most frequently used processes applied in image processing. The two fields are compared by a classic method, the analysis of typical feature descriptors best fitting convolutional networks while concentrating on the general difficulties of image processing.
I will review factors that may constitute difficulties in applying convolutional networks, factors that may justify the use of classic solutions. In order to compare the two fields I will present two challenging tasks encountered in practice, video stabilization and face identification, where we face the dilemma of which method to choose. Along with the problems that these two tasks involve I will also demonstrate solutions that are widely used.
Face identification is considered to be an extremely difficult task due to the rather restricted training dataset available. This is the reason why it is disputed whether the application of neural or classic image processing apparatuses is more effective. In my thesis I will try to respond to this challenge by comparing the solutions I implemented for solving these tasks.