Predictive coding for document layout characterization

We propose a new approach to document image layout extraction using rapid feature analysis, preclassification and predictive coding. First, a set of layout features is used to render the image profile information. The knowledge base is utilized to rule these early regions into layout labels. The regions found are given a classification tag and a degree of membership into background, text, picture and line drawing classes. A predictive coding method is used with the preclassification information to increase the confidence of each label, and to integrate the regional domain and the labels into a uniform class without any shape assumption. We have tested our technique using three different databases that comprise over 1000 document images. The results show a high degree of confidence in region separation and extraction. The main benefits include robust classification shape independency and rapid computation.