Skip to Main Content
Practically document images may have complex background in form of non-uniform illumination (shading) or an image in background. Such complex backgrounds result poor binarization causing character recognition errors. If such images are transmitted over a noisy analog channel, they are also corrupted by white Gaussian noise that makes binarization even worse. In this paper, a denoising and binarization scheme of document images to make them suitable for OCR using discrete Curvelet transform is presented. The proposed Curvelet based method is able to remove complex image background as well as white Gaussian noise and results in a better binarized document image as compared to other conventional methods. The ability of sparse representation and edge preservation of Curvelet transform helps better in text shape preservation even in the presence of noise. The proposed method is able to remove low frequency complex backgrounds and high frequency Gaussian noise and their combinations from document images and shows better performance in such noise combination cases when compared to commercial OCR packages.