The Bag of Words(BoW) model is applied to object classification. An improved algorithm based on soft quantification is proposed. This algorithm quantifies the Scale-Invariant Feature Transform (SIFT) descriptor into several nearest visual words and performs weighting on them, which can conserve the space information. Therefore, it can avoid the information loss of eigen space caused by hard quantification. The experiments are carried out on Caltech 101 database. Experimental results show that the proposed method performs better than traditional method on image classification.