US 12,278,827 B2
Document image classifying system
Md. Rafiul Hassan, Dhahran (SA); and Muhammad Imtiaz Hossain, Dhahran (SA)
Assigned to KING FAHD UNIVERSITY OF PETROLEUM AND MINERALS, Dhahran (SA)
Filed by KING FAHD UNIVERSITY OF PETROLEUM AND MINERALS, Dhahran (SA)
Filed on Jul. 10, 2024, as Appl. No. 18/768,082.
Application 18/768,082 is a continuation of application No. 17/510,458, filed on Oct. 26, 2021, granted, now 12,095,781.
Prior Publication US 2024/0364718 A1, Oct. 31, 2024
This patent is subject to a terminal disclaimer.
Int. Cl. H04L 29/06 (2006.01); G06N 20/00 (2019.01); H04L 9/40 (2022.01)
CPC H04L 63/1416 (2013.01) [G06N 20/00 (2019.01)] 8 Claims
OG exemplary drawing
 
1. A system for classifying document images, including:
processing circuitry configured to
train a machine learning classifier, including
perform a Hidden Markov Model (HMM) for generating log-likelihood scores based on a plurality of attribute value vectors for one class and attribute value vectors for another class for page layout structure characterizing a document image, wherein the generating includes recursively computing the log-likelihood score of each state of each of the attribute value vectors, and wherein there are an insufficient number of attribute value vectors in the one class to train the machine learning classifier for the one class, wherein the document image has at least one representative block, wherein the at least one representative block characterizes the page layout structure by an attribute value vector for block features including height of the block, length of the block, area of the block, eccentricity of the block, percentage of black pixels within the block, percentage of black pixels after the application of a Run Length Smoothing Algorithm (RLSA), mean number of white-black transitions, total number of black pixels in the block after RLSA, and number of white-black transitions in the original bitmap of the block,
rank log-likelihood scores generated by the HMM,
group the plurality of attribute value vectors into a predetermined number of bins, wherein the attribute value vectors in each bin are grouped by log-likelihood scores within equal ranges,
apply a one-sided sampling technique on each bin of the predetermined number of bins in order to remove redundant and borderline attribute value vectors of the attribute value vectors of the another class in the respective bin in order to obtain a balanced training dataset between the one class and the another class in each bin, and
train the machine learning classifier using the respective balanced training dataset, and
classify document images as one of a plurality of page types using the trained machine learning classifiers based on the block features of the page layout structure.