Codebook has been shown to be an effective image representation method. In this method, discriminative local features, e.g., SIFT, are extracted from images and then pooled together. All these local features are then clustered and the centers of all the clusters form a codebook. By counting the distribution of local features on these codes, we obtain a histogram of local features as the global feature of the images. In the codebook representation, the number of the codes has an evident influence on the performance of the codebook. There have been some works dealing with this problem. However, they often determine the SVM regulation parameter by cross validation. Since the cross validation is usually computationally expensive, in this paper we investigate the effect of different SVM regulation parameters on the codebook performance. Our experiments on several image datasets provides useful observations.