AN INTELLIGENT SPARSE VISUAL FEATURE DESCRIPTORS FOR CONTENT-BASED IMAGE RETRIEVAL

Abstract

Content-based image retrieval (CBIR) techniques are used to retrieve similar images from image repositories by utilizing the visual contents of the images. From last few years, bagof- visual-words (BoVW) model is most commonly used for image retrieval and got promising results in terms of accuracy and effectiveness. However, BoVW model still has some problems, such as an image is represented as an orderless global histogram of visual words that neglects the spatial layout of the image. Spatial information is an important component that provides discriminating details for accurate retrieval of images. In this thesis, three novel approaches for image representations are presented by the selection of appropriate semantic regions of an image by constructing histograms of visual words. The standard image databases are used to determine the efficiency of proposed approaches. Following approaches are presented in this dissertation: A novel image representation is presented using the characteristics of local and global information in the form of histograms of visual words. The global information is obtained by constructing the histogram of visual words over the whole image, while the histogram of visual words for local information is constructed over the local rectangular region of the image. The local histogram represents the spatial information of salient objects. In order to verify the performance of the proposed approach, a number of experiments are conducted on the standard image databases (Corel-A, Caltech-256, and Ground truth). The results show that the proposed image representation significantly enhance the effectiveness of image retrieval. Based on the semantic similarity in an image, another image representation is proposed by constructing the histograms of visual words by splitting an image into two rectangular regions that add the spatial information to the inverted index of the BoVW based image representation. By utilizing this phenomenon of image representation, different visual words for upper and lower rectangular regions of an image are obtained for better image retrieval performance. For the verification of proposed approach, extensive experiments are conducted vii on Corel-A, and Ground truth image databases, proof the robustness of the proposed approach. In order to overcome the problems of overfitting on large dictionary sizes, lack of spatial information, and to reduce the computational cost, a new image representation based on the weighted average of triangular histograms (WATH) is also introduced. The image is divided into four triangular regions in order to incorporate the spatial information to the inverted index of the BoVW based image representation, and a histogram of visual words are computed from each triangular region. An appropriate weight is assigned to each histogram in order to eliminate the aforementioned problems. The assigned weight reduces; the size of the dictionary by reducing the non-salient visual words, and the computational cost. The proposed approach also provide the consistent performance on large dictionary sizes. The quantitative and qualitative analysis conducted on two image databases (Corel-A and Corel- 1500) shows the robustness of the proposed approach among the recent image retrieval approaches. Keywords: Content-based image retrieval (CBIR); Bag-of-visual-words (BoVW); Local and global histograms; Rectangular spatial histograms; Weighted triangular histograms.