Numerical sequence extraction in handwritten incoming mail documents

Abstract

In this communication, we propose a method for the automatic extraction of numerical fields in handwritten documents. The approach exploits the known syntactic structure of the numerical field to extract, combined with a set of contextual morphological features to find the best label to each connected component. Applying an HMM based syntactic analyzer on the overall document allows to localize/extract fields of interest. Reported results on the extraction of zip codes, phone numbers and customer codes from handwritten incoming mail documents demonstrate the interest of the proposed approach.

Multi-view hac for Semi-supervised Document Image Classification

“…For example, sorting incoming mails in companies allows to redirect an unknown document to the right department or to apply an appropriate processing depending on its document class [1] [2]. However, these softwares are not able to extract all the information on the image yet and a human operator has to define the tasks that have to be accomplished by the software depending on the document class of the image.…”

<title>A generic method for structure recognition of handwritten mail documents</title>

“…This problem encouraged people to restrict their analysis to specific goals. Thus, Koch et al 3 propose a HMM based analyzer to extract numerical fields in handwritten incoming mail documents. Their work aims at detecting fields like phone numbers, customer reference, zip code .…”

A Syntax-Directed Method for Numerical Field Extraction Using Classifier Combination

“…While arguing in the same way on all states, we get the following syntax model (figure 2): Note also that to complete the model, we have moreover to define a matrix of initial and final states. This model is described in detail in [10].…”

“…Our first approach was based on a non-parametric classifier (KNN) for discriminating the connected components using a contextual/morphological feature set [10]. Although this initial system gave encouraging results, we showed that the contextual/morphological feature set appears to be insufficient to correctly discriminate the classes involved.…”

“…In order to avoid the full page recognition, we have rather turned towards an original approach that only label the connected components as "belonging to a numerical field" or not [10]. Once the component labeled, the syntactical structure of the expected numerical sequences is used to filter the fields of interest in the document.…”