Coarse Coding For Material and Object Identification

Abstract

A new coarse-coding technique is presented for labeling image pixels and regions to match exemplars or multivariate material signatures. This multinomial classification method can be used for object cuing and tracking, as well as for material identification and image segmentation. Pixels are classified—and classification reliability can be estimated—with only single-band histograms and one pass through each image band. An example of four-class labeling illustrates the power of this two-level classification algorithm.