Unsupervised learning technique for binarization of gray scale text images

This paper suggests an unsupervised learning technique for binarization of gray scale text images based on fusion of existing approaches. The existing binarization approaches have though resolved issues of degraded contrast, varied intensity, shadow, smudges, bleed through and smear to a certain extent, yet none of the existing algorithms are capable to work out all the issues independently. However, these have also led to issues like broken characters and merged characters. Some of the standard algorithms in this field have been chosen as the core methods to produce partial binarization results. The decision on these partial binarized results has been fused using three different proposed approaches. The benchmark of performance evaluation are those set by DIBCo. The standard dataset along with the ground truth have also been taken from the same. Our binarized results produced from the above methods have given better results in contrast to DIBCo-2011 and have also surpassed the failure cases of DIBCo 2011. The Results and Analysis section comprehends the conclusion drawn.