Facial expression representation based on Gabor features has attracted more and more attention and achieved great success in facial expression recognition for some favorable attributes of Gabor wavelets such as spatial locality and orientation selectivity. A large number of Gabor features are produced with varying parameters in the position, scale and orientation of filters, which cause huge computational complexity. In some existing methods, useful discriminatory information may be lost due to down sampling Gabor features directly. To reduce the loss, a Gabor features representation method based on block statistics is proposed in this paper. In addition, Support Vector Machine is used to match the features. The effect of this method is demonstrated by template matching test on JAFFE database, and the comparative test results show that this method can yield better recognition accuracy with much fewer Gabor features as well as less CPU time of feature matching.