A Local Adaptive Region-wise Histogram Correction and Thresholding Technique for Very Poorly Illuminated Images

Abstract

The present work proposes a novel scheme for binarization of poorly illuminated images, that are often encountered in scanned collections of printed and handwritten texts. The readily available techniques such as adaptive mean thresholding, adaptive gaussian thresholding, Otsu's binarization, etc. usually fail in such situations, mostly because of lack of contrast in the images. There are several examples of poorly scanned documents, which besides exhibiting poor contrast, contain parts of texts that have similar intensity levels to that of some portions of the background. The methodology developed here is designed specifically to tackle situations like this. A novel adaptive region-wise histogram correction technique is developed that is capable of automatically enhancing the contrast of such images for the purpose of further processing. The enhanced images are then binarized using a region-wise thresholding technique that uses statistical methods to calculate the threshold values for different regions. Final result is an automatically generated clean binarized version of a very poorly illuminated text image.