The Research On Image Classification Based On The Bag Of Words Model

As the research directions for the analysis and understanding of digital image, Image classification is payed more and more attention.It has become a hot issue in the field of image processing that how to classify a large number of the digital image information fast and accurately.The Bag of Words model was applied in unsupervised image categorization area in this paper. In order to overcome the disadvantages of the model, some improvement has been done:First, Speeded Up Robust Features (SURF) has been used to extract features (image words) in this article, this article will preprocessing images. To specified the feature extraction area by Region of Interesting Features (ROIF) and the distinguish technology about foreground and background. Experiments show that the preprocessing can effectively reduce the number of weak and irrelevant features.Secondly, Exact Euclidean Locality Sensitive Hashing (E2LSH) has been used in this article.In order to reduce the randomness of E2LSH, require several clustering and using undirected graph pooling technology make several image dictionary synthesize a single dictionary. Experiments show that the representative of the E2LSH’s clustering centers has stronger.Thirdly, this article using Gibbs sampling to estimate the implicit transfer matrix by the latent Dirichlet distribution (LDA) model.And use a new interpretation method combined the maximum transition probability and transition vector similarity to obtain classification results.Finally, in order to traditional bag of words models ignore the within class image relations. In this paper, Bidirectional Matching (BM), Random Sample Consensus (RANSAC) and Perceptual Hash (pHash) have been used to find the relationship within the same class image. Experiments show that this method can accurately obtain the relationship within the same class image, to achieve a detailed classification within a class.