Modelling human limits for visual data assimilation

Summary form only given. Some aspects of the performance of the human visual system (HVS) have been quantified in psychology and neurology experiments and theories. Much of this knowledge does not extend beyond the early vision stages, and so does not provide an adequate basis for accurate models of how humans perceive particular instances of visual data, such as individual images or sequences of images. Often the various visual effects are assumed to apply independently of all other aspects of HVS behaviour, and to be of comparable importance when occurring together. Such restrictions do not allow more general modelling to be applied successfully except for very simple cases, and preclude the modelling of complex images by simple combination of the visual effects separately. Due to the high data rates associated with digital images and image sequences, leading to a growing need for higher compression rates, a more tractable system for modelling HVS response generically must be devised. This would allow compression techniques to be targeted to discard more information which is redundant in terms of visual assimilation. The use of a visual attention paradigm can be considered as the basis for such a system. This paper present an overview of the proposed paradigm for visual attention and importance modelling, and use this to conjecture some limitations on the image structure and quality, rate of change and rate of information delivery needed for typical visual data delivery services. Synthesized examples of scenes coded using this model are given and linkages to related work in compression and quality assessment are described.