Putting knowledge into a visual shape representation

摘要

This paper shows how a representation for visual shape can be formulated to employ knowledge about the geometrical structures common within specific shape domains. In order to support a wide variety of later visual processing tasks, we seek representations making explicit many geometric properties and spatial relationships redundantly and at many levels of abstraction. We offer two specific computational tools: (1) By maintaining shape tokens on a Scale-Space Blackboard, information about the relative locations and sizes of shape fragments such as contours and regions can be manipulated symbolically, while the pictorial organization inherent to a shape's spatial geometry is preserved. (2) Through the device of dimensionality-reduction, configurations of shape tokens can be interpreted in terms of their membership within deformation classes; this provides leverage in distinguishing shapes on the basis of subtle variations reflecting deformations in their forms. Using these tools, knowledge in a shape representation resides in a vocabulary of shape descriptors naming constellations of shape tokens in the Scale-Space Blackboard. The approach is illustrated through a computer implementation of a hierarchical shape vocabulary designed to offer flexibility in supporting important aspects of shape recognition and shape comparison in the two-dimensional shape domain of the dorsal fins of fishes.