Learning codeword characteristics for image retrieval using very high dimensional bag-of-words representation

Abstract

In this thesis, we address three problems of the tf-idf weighting, the hierarchical scoring and intra-class key-words in the image retrieval using very large dimensional Bag-of-Words (BoW) representation. The tf-idf weighting method, which is a commonly used codebook weighting scheme, is usually understood to improve retrieval performance however its degree is not too significant. Rather, it sometimes brings a problem of worsening precision. The hierarchical scoring, which is commonly used in hierarchical codebook, the precision improvement differs depending on the dataset as the number of levels that are considered in scoring gets larger. Intra-class key-words, which represent their class most well, have not taken into consideration in BoW representation based image retrieval because of its too high dimensionality. Despite different classes have different key-word, only a same weight or standard is applied to every image classes. To overcome these three different problems, we suggest a new codewords weighting method preserving the independence model of BoWs representation that codewords occur independently in one image. In the problems of tf-idf weighting and the hierarchical scoring, the proposed method only focuses on improving the algorithms without using any extra cue besides two types of signature, document frequency used for tf-idf weighting and the level for hierarchical scoring. Since the document frequency and the level are related with inter-class discriminability, we define the two values as key-signatures. In the problem of considering intra-class key-words, the proposed method gives relevant weight to codewords according to its statistical appearance within a class. Since an intra-class variance of frequencies of a codeword is related with its intra-class importance, we define that kind of value as another key-signature. We also define a function, called Weight Mapping Function (WMF), that maps a weight value from a key-signature. In order to obt...