Perceptual image coding using a cortical snapshot model of human vision.
Michael Horowitz, David L. Neuhoff · Deep Blue (University of Michigan) · 1998
We introduce a two stage image coding algorithm called Perceptual Pruning (PP). In the first stage, a perceptually lossless reproduction is produced using an adapted basis algorithm called matching pursuit. In the second stage, we prune components of the matching pursuit reproduction that are deemed perceptually irrelevant by a newly developed indistinguishability criterion. We demonstrate that PP is capable of producing reproductions with higher subjective quality than reproductions produced by several state-of-the-art image coding algorithms at various encoding rates. In addition, the PP coder is the first image coding algorithm to fully integrate a sensor based model of the HVS into an adapted basis image coding technique. At the heart of the PP coder, lies a novel indistinguishability criterion based on a maximum likelihood rule for how the human visual system (HVS) distinguishes two images. Our criterion is novel in that it combines inputs of different types at different spatial locations. Existing discrimination functions combine inputs of different types all located at one location in space. We show that this spatial pooling leads to an indistinguishability criterion that more accurately predicts sensitivity characteristics of the HVS than when no spatial pooling is used. The indistinguishability criterion is developed around the notion of a cortical snapshot, a term we coined to describe the input to the striate cortex of the brain at one point of fixation. As the eye subconsciously scans through an image in some fashion, the cortical snapshots at each point of fixation are passed on to a higher level of the cortex, where they are collated into the perception of the complete image. Specifically, the indistinguishability criterion is based on a simple sensor-based model of the small subset of the cortical snapshot found in a circular region with a diameter of approximately three degrees, centered at the point of fixation.