Conventional Binarization methods try to obtain optimal results based on the single image only. They make distinct diversity of binarization quality sometimes even for images of the same documents. Using a binarization evaluation and feedback mechanism, this paper proposed a learning-based binarization method which can improve the binarization of same-type document, especially in the quality stability. It has a learning and a performing binarization stage. Learning stage obtains knowledge for binarization evaluation and optimization. In performing stage, the evaluation of binarization result is fed back to binarization in order to adjust binarization parameters, which will improve the binarization. Experiments validate the improvement.