A Model for Facial Activity Recognition using Metarepresentation: a Concept release_5ja7bhnnanhijpe4iomrevwlei

Abstract

Recognition of the facial visual properties (physiognomy) and its static and dynamic behavioral patterns (action units) has proved to be an important part in many multimedia retrieval and analysis applications. Apart from the previous studies, where methods to extract part of the action units from an image or video have been developed, in this ongoing research project we work on a model for more accurate and detailed facial activity semantic description adaptable to new behavioral patterns and real conditions. In this paper, we address challenges of building this model and suggest its basic multilevel concept. On the low level, we propose using wavelet-based multiresolution representation of video data. On the middle level, several multiclass classifiers are being examined for the purpose of attribute learning, and a custom multiple metric is provided. On the high level, facial elements, behavioral patterns and their attributes can be connected and further extended using the ontologically-compliant architecture of this model. On the abstraction layer, all three levels of this model are seamlessly integrated via graph-based hierarchies of metavertices, metaedges and their mappings. Having this structure, the proposed model can be trained and employed to solve the problems of human behavior retrieval and human-computer multimodal interaction more efficiently. Current results, however, reveal that to be reliable, this model requires further research studies and their comprehensive experimental evaluation.
In text/plain format

Archived Files and Locations