A simple text detection in document images using classification-based techniques

Text regions can be useful to computer vision applications. It can be used to label and train automatic layout learning systems or to detect and locate the title, keywords, subheadings, paragraphs and image regions in images. This work proposes a technique to separate text regions from image documents. Images are divided into small non-overlapping windows. Textural features are extracted from these image windows before a classification is performed. Two refinement processes are carried out to reject misclassified windows, i.e window merging and Markov Random Files (MRFs). Window merging determine the similarity of a window and its neighbouring windows (based-on a distance-based technique). MRF examines the relationships between each window and it's neighbouring one using an energy minimization technique. The experimental results demonstrate that the refinement method is superior to the original classification without a refinement.