Media Converter with Impression Preservation Using a Neuro-Genetic Approach

Abstract

We propose a media converter framework which takes images, music, and other media as input and outputs different media while preserving the same impression for humans. The media converter is realized by combining multiple media database (DB) retrieval systems that use a common psychological impression space. Each of the media DB retrieval systems consists of a physical media feature extraction part, a physical feature space for the extracted features, an impression space where the psychological impressions of media are expressed as their coordinates, a neural network (NN), and a genetic algorithm (GA). The NN maps the coordinates of the media in the physical feature space to those in the impression space, and the GA search the coordinates of media in the physical feature space using the target coordinate in the impression space and the NN mapping. A user specifies a coordinate in the impression space that corresponds to his or her target impression. The media DB retrieval system extracts features of the images, music, or other media on the Internet or a commercial packaged media DB and stores them inside as physical feature space coordinates. Media whose impression is similar to the user-specified target impression are searched by a NN and GA. The media converter is realized by combining multiple media DB retrieval systems. A medium, medium A, is converted as follows: (1) it is mapped from a physical feature space A to an impression space by NN$_A$ and (2) medium B is searched for in a physical feature space B from the impression space using GA$_B$. Prototypes of an image DB retrieval system and a music DB retrieval system are made and evaluated for their mapping and searching performance. Finally, we make a prototype of a media converter by combining the media DB retrieval systems and show the potential of its realization.