Flexible and Robust Model Matching based on Association Graph for Form Image Understanding

Abstract

A new method of image understanding for forms based on model matching is proposed in this paper as the basis of an OCR which can read a variety of forms. The outline of this method is described as follows. First, ruled lines are extracted from the input image of a form. After that, several lines are grouped as one to be recognised as data corresponding to a sub-form. These lines and subforms are both used for understanding the form, taking into account their feature attributes and the relationships between them. Each feature in the input image of a form is expected to correspond to a feature in one of the model forms, which are described as structured features. This correspondence is represented by a node in an association graph, where an arc represents compatible correspondences established on the basis of feature relationships. The best match is found as the largest maximal clique in the association graph. Experimental results show the method is robust and effective for document images of poor quality, and also for various styles of forms.

A hierarchical method for automated identification and segmentation of forms

“…The problems with these approaches are that the feature vector does not reflect the hierarchical layout of the forms. Another type of representation tries to utilise hierarchical structure of blocks as an X-Y tree [2,14,7]. X-Y tree representation is possible only for forms with boxes and thus has limited application area as many forms have horizontal marker lines instead of boxes.…”

Table structure analysis based on cell classification and cell modification for XML document transformation

“…In this case, a text region in each data field of a table is hierarchically extracted. This subsystem also extracts table features consisting of the geometric information of a table, that of ruled lines, and that of data fields from each table region [4]. In the proposed method, a data field is described as a rectangle and is called a cell.…”

“…In the proposed method, the table having simplified cells is identified and is modified to construct the regular row structure of cells from the input table as follows. First, an isolated table is detected from the input document by ruled line extraction and grouping process [4]. Next, each cell is extracted from these ruled lines and is classified.…”

Recognition and identification of form document layouts

“…The method is found to be hard to apply to analysis of filled-in form, because it is considered to be limited to empty fields. Ishitani[10] uses a hierarchical matching strategy based on sub-graph matching, local matching stages by line matching, and the interactions between them. But the similarity measure totally depends on the number of vertical and horizontal lines.…”