A document analysis system
Jiangying Zhou · 1993
Building a fully automated document processing system remains a challenging problem for researchers. The work presented in this thesis expands the document image processing techniques in two aspects: (1) how to decompose a document page into sub-regions which represent coherent components such as text, graphics, photographs without the knowledge on the specific layout model. (2) how to describe character shapes. Using the geometry properties of background white spaces we have found a way to avoid a costly skew correction in the separation of columns in a document image. An efficient segmentation is thus achieved in the presence of severe tilt. In a study of an important special case of block classification, an interesting blend of theoretical analysis and practical experiments shows that a faster linear algorithm can be devised using properties of the across scanlines correlation to separate photographic regions from text/line-drawing regions. We proposed a hierarchical description method of curve shapes. The method is a result of the work that has been directed towards improving performance of an OCR system. As an application of the description method, we proposed a machine printed character recognition system. We demonstrated that the scheme is very powerful in both encompassing wide variations in the shape of individual characters and in discerning characters of similar shapes.