Local Contrast Based Thresholding for Document Binarization

Abstract

Global thresholding and local thresholding are two basic ways for image binarization. Global thresholding, not suitable for complex documents, may produce noise along the page borders when the intensity of an image is non-uniform. For degraded, noisy and illuminated images local thresholding is preferred. This paper presents a local thresholding scheme when the image is a collection of text, pictures, and background. It calculates the local mean and local contrast to separate the background from the foreground. Local contrast, not suitable for integral images, can be achieved by the differing local maximum, and local minimum intensity. The proposed approach doesn’t calculate standard deviation like much other local thresholding method but locally adapted with local mean and local contrast. This proposed algorithm has been tested with different types of images including ancient documents, noisy background, and pictures. The test results are compared with a global and some traditional local thresholding techniques in terms of MSE and PSNR. The comparison shows that the proposed method results improved PSNR and better extraction in OCR operation.