One of the difficulties of extracting text contained in images or videos comes from the variation of the grayscale values of the text and backgrounds. In this paper we propose a new method to normalize the contrast between text characters and backgrounds so that a trained machine learning tool can verify characters of grayscale values that have never been seen before. Experiments show that the proposed method used in training either a multilayer perceptrons or a support vector machine yields better text verification comparing with other typical contrast measures.