Method of neighbor embedding for OCR enhancement

摘要

Single-image super-resolution (SISR) is the problem of generating a high resolution image from a single low resolution image. The SISR technique known as neighbor embedding utilizes a training ensemble of pairs of low and high resolution image patches, where the patches in a given pair represent the same image region. The present invention improves upon prior neighbor embedding algorithms by offering a practical, computationally efficient method of neighbor embedding for generating a high resolution version of a low resolution image. The technique may also be applied to generate high resolution versions of low resolution text images for subsequent input into OCR engines. OCR character error rates found on the high resolution images are drastically lower than those found when OCR is applied to the original low resolution text images.

主权项

1. A method of neighbor embedding comprising the steps of:
creating a training of pairs, each having a first element, FLn,m, and a second element, H*n,m; creating a set of randomized k-d trees from said training set by using said first elements, FLn,m, as leaf nodes; receiving at least one low resolution text image, P; generating a high resolution image, T, from said at least one low resolution text image, P, using said training set and the result of step (b); and generating a text file by inputting said high resolution image, T, into an OCR engine; wherein the step of creating a training set comprises:
a. receiving a user-defined set of text documents, Sn;b. generating, for each Sn, a high resolution image, Hn;c. generating, for each Sn, a low resolution image, Ln;d. selecting a user-definable patch dimension, K, where K is a user-definable small positive integer;e. partitioning said low resolution image, Ln, into a set of K×K patches, Ln,m each having K×K pixel values;f. setting J=NK, where both N is a user-definable small positive integer;g. partitioning said high resolution image, Hn, into a set of J×J patches, Hn,m, each having J×J pixel values, such that said K×K patch Ln,m and said J×J patch Hn,m both correspond to a same region in Sn;h. generating a feature vector, FLn,m, having K×K components, for each said K×K patch, Ln,m; andi. generating a J×J high resolution patch, H*n,m, for each said J×J patch, Hn,m, by subtracting a mean pixel value of said corresponding K×K patch, Ln,m, from each of said number of pixel values in said J×J patch, Hn,m.