Document segmentation using texture analysis

This paper describes the application of a novel texture recognition strategy, which involves the assembly of n^th order co-occurrence information within a processing window. The resulting spectrum is processed by a nearest neighbour classifier, and discontinuities in the classification as the operating window scans the image are recognized as texture boundaries. The technique is being applied for the first time to segmenting newsprint, in order to locate the text regions for effective data capture. Previous methods applied to the task discussed in the paper, have used run-length encoding and block analysis, Gabor filters, or component analysis. The co-occurrence method offers significant advantages in processing time and ease of implementation