In this article, we propose a special type of decision tree, called a decision cascade, for binarizing document images. Such images are produced by cameras, resulting in varying degrees of brightness over different parts of the images. Our method decides what action to take on each part of the input image in order to obtain a satisfactory binary result. The advantage of this approach lies in its ability to learn decision rules from training data that may be labeled with multiple identities. Tests on images produced under improperly illuminated conditions show that our method yields much better visual quality and OCR performance than using a global threshold for binarization.