The existing visual dictionary construction methods need to combine several features into a vector. Then the vectors are clustered to form the dictionary. Those approaches only take the similarity of all the features into consideration but the neglect distinct roles of diverse features on the construction of the visual dictionary. In this paper

a visual dictionary construction method based on Dempster-Shafer (D-S) evidence theory is proposed. D-S evidence theory is applied to fuse different features in their similarities

which is helpful to obtain more accurate visual dictionaries. Two kinds of features are applied in this paper to subdivide the initial visual dictionary based on the Dempster-Shafer evidence theory

and similar features are clustered together better. Compared with the traditional visual dictionary generation method

our proposed method obtains better results. The experimental results on image classification using support vector machine (SVM) and Naive Bayisan (NB) classifiers show that our proposed method outperforms the K-Means based dictionary construction algorithm in terms of accuracy in visual dictionary and image classification.