Analysis-by-synthesis and example based animation with topology conserving neural nets

Introduces the method of “template-based image generation” (TBIG) which is based on self-organizing neural networks and aims at producing images and animations from still images. The main characteristics of the method are (a) storage of image data in self-organizing maps (SOMs) and (b) a dual image representation which requires the reliable automatic identification of interest points; the latter task is solved by a hierarchical analysis/synthesis system based on neural nets. Current work focuses on images of the human face. The new possibilities of storing, encoding and manipulating images are integrated in the “FaceCoder” system which is intended to work with image data in the same way as vocoders do with speech data. Sequences of expressions or articulatory movements are “transferred” from an image sequence to a still image. Applications are compact image encoding, production of phantom images for forensic use, and special effects for entertainment purposes