This paper studies a novel visual word generation method in the Bag-of-words model for object categorization. The conventional Bag-of-words algorithm represents the cluster centers as visual words, which led to the incomplete expressions of image semantic information, so an improved method for visual word generation using the soft-decision based on kernel function is proposed. First, SIFT keypoints of images are extracted. Then, after clustering SIFT keypoints, some typical SIFT keypoints are selected from a cluster by kernel density estimation using a kernel function. Finally, these selected keypoints are trained employing SVM to generate a visual word of this cluster. Experimental results show that the proposed visual word generation method enhances the expressions of image semantic information, increases the recall ratio effectively, and improves significantly the effect of object categorization.