Hidden tree markov models for document image classification

Abstract

Classification is an important problem in image document processing and is often a preliminary step towards recognition, understanding, and information extraction. In this paper, the problem is formulated in the framework of concept learning and each category corresponds to the set of image documents with similar physical structure. We propose a solution based on two algorithmic ideas. First, we obtain a structured representation of images based on labeled XY-trees (this representation informs the learner about important relationships between image sub-constituents). Second, we propose a probabilistic architecture that extends hidden Markov models for learning probability distributions defined on spaces of labeled trees. Finally, a successful application of this method to the categorization of commercial invoices is presented.

A new efficient probabilistic model for mining labeled ordered trees

“…Here we briefly describe the differences between OTMM and two other similar tree Markov models, called the Hidden Tree Markov Model (HTMM) [6] and the Probabilistic Sibling-dependent Tree Markov Model (PSTMM) [18,19].…”

“…However, directed acyclic graphs are much more complex than labeled ordered trees, and so simpler probabilistic models and more efficient learning algorithms for trees have been proposed. For example, the hidden tree Markov model (HTMM) [6] is a probabilistic model for labeled trees that handles parent-child dependencies in a given tree. However, it does not handle sibling dependencies, meaning that HTMM is not applicable to labeled ordered trees.…”

Fisher kernel for tree structured data

“…We decided instead to restrict ourselves to a more tractable class of models, the so called Hidden Tree Markov Models (HTMMs). This architecture was first suggested in [7] and developed in details in [6] for document processing. Please refer to those papers for a comprehensive description.…”

“…This, in a Bayesian Networks (and thus also in HTMMs), is performed by first computing the expected sufficient statistics for the parameters and then updating parameters with the normalized sufficient statistics. For an explicit equation formulation refer to [6].…”

“…Actually, mainly for computational reasons, as we will explain in the subsequent sections, in the experiments we work with Hidden Tree Markov Models (HTMMs) [6]. They represent a subset of HRMs to deal with trees and we modify them adding Gaussian nodes to represent the target values.…”