An Analysis of Image Binarization Techniques for Natural Scene Images

Text extraction from natural scene images is an emerging field in computer graphics. Extracted text contains important information that can be used for various purpose like vehicle number plate detection to identify the vehicle, to provide information of surrounding to visually impaired persons, preservation of information of historical documents etc. Binarization is a key process in text extraction process. It is challenging take in case of natural scene images due to uneven lighting conditions, complex background and unpredicted text size, color and layout. Three well known binarization techniques namely Otsu’s, Niblack’s and Sauvola’s binarization techniques are test on natural scene images. We found that, Sauvola’s algorithm can achieve better performance than Niblack’s. In most of cases Sauvola and Niblack gave good results as compare to Otsu’s method. Otsu binarization technique is good for uniform background. Window based Niblack’s and Sauvola’s methods are useful to find local threshold to binrize natural scene images.