Authors: [tex2html_wrap4112]M. Kelly, M.D. Levine
Investigator username: levine
Subcategory: computer vision
The visual capabilities of higher invertebrates have been studied in order to gain insight into how shape information may be represented in these organisms. Some cephalopods are able to make distinctions between different objects based solely upon the projected image of two-dimensional bounding contours. These distinctions do not depend upon the absolute size of the retinal projection, nor are they affected by certain types of deformations or variations in object shape. Our studies have given us some insight into how relatively crude information may be extracted in order to provide approximate representations for two-dimensional shape. Within the larger context of shape recognition, the flexibility afforded by such representations could provide significant improvements over traditional computer-based vision systems.
As part of this study, we have considered a means for detecting regions enclosed by sets of planar curves. Symmetric relationships between the enclosing contours can be identified, thereby providing a compact way to represent these regions. Two distinct types of symmetric enclosure are identified. A perceptually meaningful distinction is made between shape components possessing a round or blob-like structure versus those which have more of an axial limb structure.
A practical computer vision implementation is being developed whereby symmetrically enclosed points are extracted directly from an image edge map. A set of multiscale edge operators are first applied to an image. Annular operators having a range of scales which match those of the edge operators are then used to identify particular edge groupings. One of the two symmetric enclosure types may be associated with any given annular operator and used as a label. The set of labelled annular centers provides a means for representing two-dimensional shape contours. Symmetrically enclosed limb points which are adjacent either spatially or across two neighboring scales may be linked to indicate the presence of an axial structure. A shape representation may then be constructed as a connected graph in which nodes correspond to individual shape entities (blobs or limbs) and the connectivity provides spatial position and relative scale information. Using multiscale sampling, it is possible to extract representations even from complex shapes possessing internal structure (texture), or in cases when the contour is highly fragmented. The computational methods used here possess inherent parallelism at all stages of operation. This gives rise to the possibility for efficient implementations having fast image processing rates.