In this paper, we propose a novel binarization method for document images produced by cameras. Such images often have varying degrees of brightness and require more careful treatment than merely applying a statistical method to obtain a threshold value. To resolve the problem, the proposed method divides an image into several regions and decides how to binarize each region. The decision rules are derived from a learning process that takes training images as input. Tests on images produced under normal and inadequate illumination conditions show that our method yields better visual quality and better OCR performance than three global binarization methods and four locally adaptive binarization methods.