Region based adaptive binarization for optical character recognition purposes

Optical Character Recognition (OCR) applications usually require the use of uniformly illuminated images which can be obtained using the flatbed scanners. However, rapid development of mobile technologies causes the growing popularity of document images captured by built-in cameras being the integral parts of modern mobile phones and tablets. Many companies and administration offices accept not only the scanned documents but also high resolution photos which can be enough e.g. for insurance purposes. Unfortunately such images can be unevenly illuminated causing some problems for the OCR applications used for text recognition especially if the QR, Aztec or some other popular 2D codes are not present. Proper text recognition from camera images requires image preprocessing including its binarization which cannot be conducted using typical global thresholding due to the presence of local intensity changes. On the other hand the use of pixel-based adaptive methods is time-consuming and not always leads to satisfactory results. To fill this gap and balance the recognition accuracy and high processing speed a region based approach to image binarization is proposed in this paper being an extension of well-known Niblack thresholding algorithm.