Research On Bag Of Visual Words Based Image Classification

With the rapid development of the Internet and multimedia technology, the amount ofmultimedia data including images and videos increases with an explosive speed. To organize,manage and retrieve images accurately and efficiently, the computer is required to understandthe content of images. Image classification is a key way to solve image understanding problems,and plays an important role in the development of multimedia retrieval technology. Currently,Bag of Visual Words (BoVW) and Support Vector Machine (SVM) have been becoming themainstream in the field of image classification.This thesis mainly researches on image classification technology based on BoVW. Tosolve the problems existed in BoVW, the corresponding solutions are proposed. The maincontributions are listed in three aspects as follows:(1) Aiming to the synonymy and polysemy of visual words, Locality Sensitive Hashing isintroduced into high-dimension vector clustering, and a randomized aggregating visualdictionary model is proposed. Firstly, instead of K-Means algorithm, Exact Euclidean LocalitySensitive Hashing (E2LSH) is used to cluster SIFT features to generate the randomizeddictionary, reducing the synonymy and polysemy of visual words. Then, considering therandomness of E2LSH clustering, a group of randomized dictionaries is aggregated viaclustering aggregation techniques, resulting in the ultimate randomized aggregating visualdictionary. Experimental results show that, this novel model improves the efficiency ofconstructing the visual dictionary, and the synonymy and polysemy of visual words are greatlyreduced. As a result, the discrimination and representation of the visual dictionary are enhanced.(2) To overcome the lack of spatial information in visual vocabulary features, a spatialpyramid co-occurrence matrix based image classification approach is presented. Firstly,inspired by the Spatial Pyramid Matching Kernel, the image is partitioned into groups of imageregions to describe the absolute location information of visual words. Secondly, acorresponding visual words co-occurrence matrix is created for each image region to describethe localized relative location information between visual words. Then, capturing the spatialinformation of visual words, all visual words co-occurrence matrixes together are combinedwith the visual vocabulary histogram, to generate the ultimate visual vocabulary feature. Finally,the SVM is used to accomplish the image classification. Experimental results demonstrate thatthis approach introduce the absolute and localized relative location information of visual wordsto visual vocabulary features successfully, and improves the image classification performance. (3) As for the correlation measure between visual words’ semantic meanings, an imageclassification approach using spatial context homoionym based soft-assignment is put forwards.Firstly, spatial context information of visual words is exploited to represent their semanticmeanings, and the similarity between spatial context information is used to measure thesynonymity between visual words. Secondly, according to the similarity between visual words,a spatial context homoionym table is created. Then, a soft-assignment scheme is implementedto realize the assignment of SIFT features to multiple visual words which own similar semanticmeanings. Finally, the SVM is used to accomplish the image classification. Experimentalresults demonstrate that, this approach effectively overcomes the shortcomings of traditionalsoft-assignment-based ones, and improves the image classification performance.