Learning Binary Shapes as Compression and Its Cellular Implementation

摘要

We have presented in this paper a methodology for fast shape learning andrecognition that fits the organization of vision systems based on programmable artificial retinas. The emphasis has been put on the general foundations and trade-offs rather than on the design of operational algorithms. There are obvious relationships with existing techniques of 'shape decomposition into meaningful parts' as well as decomposition into particular structuring elements. From this point of view, compression stands out as a unifying concept. As far as retinal adaptive target tracking is concerned, the tree representation we have come up with now has to be improved to become robust with regard to shape deformation, occlusion and rotation.