Syntactic and Semantic Image Representations for Computer Vision

Computer vision requires the processing of images at various levels of abstraction. This thesis explores two image representations for vision which can be classified as "syntactic and "semantic respectively. As an exploration into syntactic (signal level) representations, an image coding technique based on the statistical relationship between subbands of the wavelet transform is explored. A sample implementation is described, which shows how this technique can be integrated into standard vector quantization coding schemes. A formulation for the recovery of rigid and non-rigid motion from optical flow is presented as an exploration of semantic (content based) image compression. This framework is based on the dynamic behavior of deformable models, and allows for closed form solution. This technique has applications for image understanding and the interpretation of visual motion. Thesis Supervisor: Alex P. Pentland Title: Associate Professor Computer Information and Design Technology