MPEG-based Scene Class Recognition

摘要

From the viewpoints of engineering, it seems reasonable to suppose finite classes in scenes that are to be recognized. First, the integrated scene recognition scheme which has the interaction capabilities between features and semantics is proposed. On the basis of this concept, the idea of scene class is investigated. Secondly, MPEG-based cross-learning tasks are performed between five representative scenes in four subjective scene classes (SSC). Thirdly, the automatic generation of computational scene classes (CSC) is experimented. At this time, although the machine-generated CSC has not shown sufficient conformance with the human-defined SSC, it was identified that the CSC could be related to SSC under some practical constraints.