Feature extraction and use with a probability density function (pdf) divergence metric

Abstract

An image of real world is processed to identify blocks as candidates to be recognized. Each block is subdivided into sub-blocks, and each sub-block is traversed to obtain counts, in a group for each sub-block. Each count in the group is either of presence of transitions between intensity values of pixels or of absence of transition between intensity values of pixels. Hence, each pixel in a sub-block contributes to at least one of the counts in each group. The counts in a group for a sub-block are normalized, based at least on a total number of pixels in the sub-block. Vector(s) for each sub-block including such normalized counts may be compared with multiple predetermined vectors of corresponding symbols in a set, using any metric of divergence between probability density functions (e.g. Jensen-Shannon divergence metric). Whichever symbol has a predetermined vector that most closely matches the vector(s) is identified and stored.