Nachricht

Am Mittwoch, den 28. Oktober 2015, hält Herr Siddharth Mittal um 16:30 Uhr im Raum P 1.4.17 einen Vortrag über seine Masterarbeit mit dem Titel:

Hierarchical Growing Neural Gas for Artificial Visual Attention

Abstract:

Visual attention research is the field of science which deals with the study of cognitive operations that help the human brain to filter relevant from irrelevant visual information. This capability of the human brain to determine relevant information so effortlessly inspires to improve the efficiency of artificial intelligent systems by employing visual attention models. Among several existing approaches, a novel approach, the Growing Neural Gas (GNG) algorithm, can be use to segment visual entities that constitute the basis for an artificial attention model. The GNG algorithm can provide information about the underlying structures in the given visual scene which can be used to construct the saliency map of that visual scene. Once the saliency map is generated, it can be used to guide visual attention. The shortcoming of GNG is that it starts with fixed set of parameters to learn the topological structure which can not change dynamically during the calculation process. As, the granularity with which the saliency should be calculated is not known because their is no prior knowledge about the given visual scene available beforehand. Therefore, the GNG algorithm has to be applied with generalized parameters which results in the wastage of resources in several instances. To deal with this shortcoming of the GNG algorithm, the goal of this thesis is to implement the GNG algorithm in a hierarchical fashion. This novel approach calculates the saliency map with different parameter settings at each layer of the GNG hierarchy to generate a final saliency map for the calculation of visual attention. Therefore, this hierarchical implementation of the GNG algorithm can be started with basic set of parameters and the proposed algorithm will itself take care of the extent to which the hierarchy should be build to calculate saliency maps for visual attention.