Research On Image Classification Based On Bag-of-words

In recent years, with the rapid development of the Internet and multimedia technology, digital resources have experienced an exponential growth state. Moreover, as an important part of digital resources, digital images are related to many aspects of our daily life. The problem of how to classify digital images rapidly and accurately has become a research focus. The Bag-of-Words model as an image classification method which based on local features has achieved great success in image classification. Considering the merits of simple and effective, a support vector machine classifier is widely used in image classification. In this paper, image classification based on Bag-of-Words model is studied:First of all, in order to overcome the defects of K-means algorithm in Bag-of-Words model, such as high complexity and treating all feature points equally within the clustering process, we proposed a K-means algorithm based on simple random sampling and feature weighting(SRW K-means algorithm). The main idea of SRW K-means algorithm is to add weights and constraints(within-cluster dispersion and global-data dispersion) to some features which obtained by simple random sampling, so that every cluster has the largest tightness and maximum distance between the cluster centers. SRW K-means algorithm reduces computational complexity, as well as improves the performance of clustering algorithm. We conduct experiments with SRW K-means algorithm and other four kinds of clustering algorithms on three commonly used benchmark datasets(VOC2007, UIUCsport and Catltech-101), and the results show that the Bag-of-Words model which used SRW K-means algorithm has the best performance, and the mean average precisions are increased by 15.32%, 10.37% and 14.51% respectively on the basis of the K-means algorithm.Secondly, aiming at the problem that high dimension of the feature space caused by kernel functions' feature mapping in a support vector machine classifier, we proposed a Dirichlet probability distribution kernel based on the whitening transformation(DPWT kernel function). The main idea of DPWT kernel function is to map feature vectors of the original feature space into whitening vectors of another feature space, and we can achieve the purpose of classifying original feature vectors by classify whitening vectors. DPWT kernel function makes the mapping feature space maintain the same dimension as the original feature space, as well as effectively eliminate the correlation between the dimensions and reduce the redundancies of the data. We conduct experiments with DPWT kernel function and other five kinds of kernel functions on the three datasets, and the experimental results show that the DPWT method has the best performance, and the mean average precisions are increased by 17.11%, 14.33% and 23.44% respectively on the basis of the Linear kernel function.At last, in order to further improve the accuracy of image classification, we conduct third experiments, and the results show that the proposed method which combines SRW K-means algorithm and DPWT kernel function together performs better than the Linear+Kmeans method, and the mean average precisions are increased by 21.95%, 15.3% and 28.44% respectively.