Bag-of-words representation has shown to be a powerful technique for image classification. In this paper, we propose a new approach to discover the discriminability of each visual word (image feature) in the codebook for each image category. A general linear model (GLM) is employed to construct new histograms of the images which are the basis for image classification. We also discuss the relations between our approach and boosting approaches and non-negative matrix factorization (NMF).