Comparative study of thresholding techniques for gray-level document image binarization

Abstract

Binarization is difficult for document images with poor contrast or illumination, intensive noise and sources type-related degradation. A new technique based on local thresholding is described in this paper. The idea of our technique is to update locally the threshold value whenever the Laplacian sign of the input image changes along the raster-scan line. The Differential of Gaussian (DoG) is used to define the sign image. This technique is tested with many images including different types of document components and degradations. The results are compared with a number of well-known techniques. It was shown that the proposed technique outperformed all other techniques studied.