| Summary: | This dissertation describes a model of the visual cortex of the cat. The model has been applied to some of the problems faced by contemporary computer vision systems. The model goes beyond previous models of visual cortex in that it models both the anatomy of visual cortex and the ability of individual cells in visual cortex to learn. The model is based on the hypothesis that image processing in the cat's visual system consists of three levels: the retinothalamic, the primary and secondary cortical (Area 17, 18, and 19), and the associative. The retinothalamic system is modeled by using operators modeling different types of retinal ganglion cells (X, Y, and W). Cells in Areas 17, 18, and 19 are modeled using CAMs, which are models of cortical cells having associative and plastic properties. CAMs modeling Areas 17, 18 and 19 use unsupervised learning to form primitives for segmenting preprocessed images on the basis of edges, moving edges, and texture. Only one associative area, Area 21, is modeled. The model of Area 21 receives input from the models of Areas 17 and 19 via the model of the Lateral Pulvinar, which transforms the segmentations into geometrical features on the basis of the two-dimensional regions. The model of Area 21 uses supervised learning to form pattern classes which are specific and hence useful to a particular domain (environment). The domains used to test the model are Roman text, Japanese text, digitized photographs of house scenes, and examples of various textures. Experiments demonstrate that the model is relevant to computer vision research because it presents a method of solving the problem of domain-specific knowledge in computer vision systems. The model also demonstrates that many techniques for computer vision systems are suggested by the anatomy and physiology of the cat's visual system.
|
|---|
