Document Type
Dissertation
Date of Award
Spring 5-31-2014
Degree Name
Doctor of Philosophy in Computing Sciences - (Ph.D.)
Department
Computer Science
First Advisor
Chengjun Liu
Second Advisor
James Geller
Third Advisor
Durgamadhab Misra
Fourth Advisor
David Nassimi
Fifth Advisor
Vincent Oria
Abstract
Content-based image search, classification and retrieval is an active and important research area due to its broad applications as well as the complexity of the problem. Understanding the semantics and contents of images for recognition remains one of the most difficult and prevailing problems in the machine intelligence and computer vision community. With large variations in size, pose, illumination and occlusions, image classification is a very challenging task. A good classification framework should address the key issues of discriminatory feature extraction as well as efficient and accurate classification. Towards that end, this dissertation focuses on exploring new image descriptors by incorporating cues from the human visual system, and integrating local, texture, shape as well as color information to construct robust and effective feature representations for advancing content-based image search and classification.
Based on the Gabor wavelet transformation, whose kernels are similar to the 2D receptive field profiles of the mammalian cortical simple cells, a series of new image descriptors is developed. Specifically, first, a new color Gabor-HOG (GHOG) descriptor is introduced by concatenating the Histograms of Oriented Gradients (HOG) of the component images produced by applying Gabor filters in multiple scales and orientations to encode shape information. Second, the GHOG descriptor is analyzed in six different color spaces and grayscale to propose different color GHOG descriptors, which are further combined to present a new Fused Color GHOG (FC-GHOG) descriptor. Third, a novel GaborPHOG (GPHOG) descriptor is proposed which improves upon the Pyramid Histograms of Oriented Gradients (PHOG) descriptor, and subsequently a new FC-GPHOG descriptor is constructed by combining the multiple color GPHOG descriptors and employing the Principal Component Analysis (PCA). Next, the Gabor-LBP (GLBP) is derived by accumulating the Local Binary Patterns (LBP) histograms of the local Gabor filtered images to encode texture and local information of an image. Furthermore, a novel Gabor-LBPPHOG (GLP) image descriptor is proposed which integrates the GLBP and the GPHOG descriptors as a feature set and an innovative Fused Color Gabor-LBP-PHOG (FC-GLP) is constructed by fusing the GLP from multiple color spaces. Subsequently, The GLBP and the GHOG descriptors are then combined to produce the Gabor-LBP-HOG (GLH) feature vector which performs well on different object and scene image categories. The six color GLH vectors are further concatenated to form the Fused Color GLH (FC-GLH) descriptor. Finally, the Wigner based Local Binary Patterns (WLBP) descriptor is proposed that combines multi-neighborhood LBP, Pseudo-Wigner distribution of images and the popular bag of words model to effectively classify scene images.
To assess the feasibility of the proposed new image descriptors, two classification methods are used: one method applies the PCA and the Enhanced Fisher Model (EFM) for feature extraction and the nearest neighbor rule for classification, while the other method employs the Support Vector Machine (SVM). The classification performance of the proposed descriptors is tested on several publicly available popular image datasets. The experimental results show that the proposed new image descriptors achieve image search and classification results better than or at par with 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 Context Aware Topic Model (CA-TM), the Hierarchical Matching Pursuit (HMP), the Kernel Spatial Pyramid Matching (KSPM), the SIFT Sparse-coded Spatial Pyramid Matching (Sc-SPM), the Kernel Codebook (KC) and the LBP.
Recommended Citation
Sinha, Atreyee, "Investigation of new feature descriptors for image search and classification" (2014). Dissertations. 167.
https://digitalcommons.njit.edu/dissertations/167
