Date of Award
Doctor of Philosophy in Computing Sciences - (Ph.D.)
Content-based image classification, search and retrieval is a rapidly-expanding research area. With the advent of inexpensive digital cameras, cheap data storage, fast computing speeds and ever-increasing data transfer rates, millions of images are stored and shared over the Internet every day. This necessitates the development of systems that can classify these images into various categories without human intervention and on being presented a query image, can identify its contents in order to retrieve similar images.
Towards that end, this dissertation focuses on investigating novel image descriptors based on texture, shape, color, and local information for advancing content-based image search. Specifically, first, a new color multi-mask Local Binary Patterns (mLBP) descriptor is presented to improve upon the traditional Local Binary Patterns (LBP) texture descriptor for better image classification performance. Second, the mLBP descriptors from different color spaces are fused to form the Color LBP Fusion (CLF) and Color Grayscale LBP Fusion (CGLF) descriptors that further improve image classification performance. Third, a new HaarHOG descriptor, which integrates the Haar wavelet transform and the Histograms of Oriented Gradients (HOG), is presented for extracting both shape and local information for image classification. Next, a novel three Dimensional Local Binary Patterns (3D-LBP) descriptor is proposed for color images by encoding both color and texture information for image search. Furthermore, the novel 3DLH and 3DLH-fusion descriptors are proposed, which combine the HaarHOG and the 3D-LBP descriptors by means of Principal Component Analysis (PCA) and are able to improve upon the individual HaarHOG and 3D-LBP descriptors for image search. Subsequently, the innovative H-descriptor, and the H-fusion descriptor are presented that improve upon the 3DLH descriptor. Finally, the innovative Bag of Words-LBP (BoWL) descriptor is introduced that combines the idea of LBP with a bag-of-words representation to further improve image classification performance.
To assess the feasibility of the proposed new image descriptors, two classification frameworks are used. In one, the PCA and the Enhanced Fisher Model (EFM) are applied for feature extraction and the nearest neighbor classification rule for classification. In the other, a Support Vector Machine (SVM) is used for classification. The classification performance is tested on several widely used and publicly available image datasets. The experimental results show that the proposed new image descriptors achieve an image classification performance better than or comparable to other popular image descriptors, such as the Scale Invariant Feature Transform (SIFT), the Pyramid Histograms of visual Words (PHOW), the Pyramid Histograms of Oriented Gradients (PHOG), the Spatial Envelope (SE), the Color SIFT four Concentric Circles (C4CC), the Object Bank (OB), the Hierarchical Matching Pursuit (HMP), the Kernel Spatial Pyramid Matching (KSPM), the SIFT Sparse-coded Spatial Pyramid Matching (ScSPM), the Kernel Codebook (KC) and the LBP.
Banerji, Sugata, "Novel color and local image descriptors for content-based image search" (2013). Dissertations. 358.