Qualitative Representations for Recognition

This thesis describes a representation for objects and scenes that is stable against variations in image intensity caused by illumination changes and tolerant to image degradations such as sensor noise. The representation, called a ratio-template, uses low-resolution ordinal contrast relationships as its matching primitives. The choice of these primitives was inspired not only by considerations of computational simplicity and robustness, but also by current knowledge of the early stages of visual processing in the primate brain. The resulting representation is biologically plausible, although there is currently no evidence to suggest that the representation is actually used by the primate visual system. Constructed manually at first, the ratio-template can be learned automatically from a set of examples. Two applications—face detection and scene indexing—are described. The ratio-template achieves detection rates higher than 90% and can process a 320× 280 pixel image in 2.6 seconds at multiple scales. Thesis Supervisor: Pawan Sinha Title: Assistant Professor