An improved scene text and document image binarization scheme

Identification of text portions have a crucial impact on intelligent transport systems, document image processing, robotics and content based image retrieval systems. So, an accurate text identification method is necessary for text based scene image processing tasks such as OCR. Scene text image binarization plays an important role in any text identification algorithm and hence in the OCR performance. In this work a novel approach to natural scene text image binarization by tracking the text boundary based on edge and gray level variance information. Further, broken boundaries are linked to construct the complete boundary map. Here, an adaptive threshold is determined based on boundary edge information to binarize the image effectively. Compared to other well known binarization methods, our method has been proved more effective in cases where the natural scene images have low contrast, low resolution, non-uniform illumination and noise. Our experiments are conducted on the datasets of ICDAR 2003 Robust Reading Competition, ICDAR 2011 Born Digital Dataset, Street View Text (SVT) Dataset, DIBCO dataset and our laboratory made Bangla Dataset. The experimental results are satisfactory.