In this paper, we present a method to improve video OCR with multiple character hypotheses. The text regions in video need to be binarized before work as the input of current OCR system. Tranditional binarization do not use any structural information about the text. Based on a certain statistic model, we define a binarization method, which is called observation function, that should satisfy a certain condition. We then present a method to construct an observation function by computing binarization results according to multiple hypotheses of characters obtained by an OCR system..