Towards a theory of model-based image coding

In model-based image coding, low bit rates are achieved by analysing a scene into moving objects and synthesising models which approximate them. The study reported is concerned with understanding the relationships between motion, shape and texture information in a model-based predictive coding system. The purpose of the study was to aid concomitant experimental work which sought to apportion bits in an optimal way between the three output bit streams; at present, due to model failure in only 4% of the image area, conditional texture updating information typically accounts for 80% of the total bit rate. The question therefore is whether model-based predictions can be improved without significant increases in the shape and motion information