Image Classification By Scalable Vocabulary Tree Based On Feature Fusion And Clustering Optimization

With the rapid development of Internet technology,image classification technology has been developed rapidly.the more popular of the image classification model are the Bag-of-Words(BOVW)and Scalable-Vocabulary-Tree(SVT),which achieve the image classification by quantifying image features into visual words,so they have a good classification effect.However,there are also some problems such as weak expression ability of features,unstable clustering effect and unclear characterization of features.Therefore,this paper focuses on the deficiencies of the BOVW and SVT,and carries out the following work:(1)In order to improve the feature expression ability of the feature extraction module of the BOVW,this paper proposes the spatial pyramid BOVW based on multi-feature fusion.Firstly,multi-scale information is added to DenseSIFT features by wavelet decomposition technology,and in feature encoding,spatial pyramid model is used to improve the spatial position information between features,and then color information is fused to improve image expression ability.Finally,this paper carrys out comparative experiments on the standard image set with different features and different number of words,which verify that the model has a higher classification accuracy.(2)Aiming at the problem that traditional K-means algorithm relies too much on initial clustering center and has one-to-one hardening points,Scalable Vocabulary Tree based on KSVD_MD is proposed.Firstly,the maximum distance method is used to select the initial clustering center,and K-SVD sparse coding is adopted for clustering.The features are divided into multiple categories to reduce the error caused by a single feature and improve the classification accuracy.Then,this paper introduces the steps of the SVT's construction.Finally,experiments on the standard image set show that the algorithm in this chapter can effectively improve the classification accuracy by comparing with other algorithms,and which was 17% higher than the traditional vocabulary tree.(3)The multi-feature fusion algorithm was combined with KSVD_MD in SVT,and the model was applied to the classification of bottle trademarks.As for the special image of trademark,the similarity between the images is extremely high due to the relationship between the first and second generations.In order to make full use of the information in the image and improve the accuracy of classification,Firstly,the improved vocabulary tree was used for the classification experiment of bottle trademark,but the classification accuracy could only reach 95.5% in the end.And then,on the basis of the SVT,combined with OCR character recognition technology to reorder,the recognition accuracy of 99.5% was finally achieved,and the robustness of the classification model was verified from multiple perspectives of rotation,noise and occlusion.Multi-feature fusion mainly combines the improved multi-scale DenseSIFT with color information to obtain the weighted features.The KSVD_MD algorithm uses the maximum distance method to determine the initial clustering center to increase the clustering stability on the basis of K-SVD.Finally,a vocabulary tree is constructed by combining multi-feature fusion with KSVD_MD,and a good classification effect is obtained on the bottle trademark data set.