Abstract: Bag-of-words image representation has shown to be a powerful technique for image classification，which creates a signature for each image using visual words（image features） in the pre-defined codebook.This paper proposes a new statistical approach to discover the discriminability of each visual word for each image category first，and then a general linear model（GLM） is employed to combine these visual words to construct new signatures of the images.Experimental results show that the approach can enhance the discriminability of each signature and thus improve the performance of image classification.