The various lighting intensity in a document image sauses diffculty to threhold the image. The conventional statistic approach is not robust to solve such a problem. There should be different threhold value for each part of the image. The threshold value of each image part can be looked as classification problem. In such a classification problem, it is needed to find the best features. This paper propose a new approach of how to use grammatical evolution to extract those features. In the proposed method, the goodness of each feature is calculated independently. The usage of the new features produce a very good result, since there are only 5 miss-classification of 45 cases