A novel feature descriptors based on orthogonal polynomial and improved local binary pattern operators for texture classification in image databases

Abstract

Texture is a fundamental property of natural images which is of much importance in the fields of computer vision and computer graphics. Texture analysis is a kind of image analysis producing measurements of the texture. These measurements may be of low-level, such as statistics of local appearance or a result of higher level processing, such as segmentation of an image into different regions or the class of the texture present in an image. The main objective of this thesis is to provide feature extraction schemes for identifying texture categories. Two frameworks for classifying the textures have been proposed. The first framework uses a complete set of orthogonal polynomial operators for texture description including five novel methods that have been proposed. The second approach extracts statistical features from the operator responses on texture images and suggests a novel feature extraction technique. The third work attempts to find an optimal orthogonal operator set for texture classification. The ability of the operators to exhibit texture characteristics has been analyzed and justified using an energy measure. A new texture coding scheme for characterizing the textures from a representation of energy matrix has been proposed, which forms the content of the proposed fifth approach. The second framework proposes two feature descriptors for texture classification that are based on Local Binary Pattern operator. A multi level Local Binary operator has been developed to assess the distribution of micro primitives at different levels and a novel Local Ternary Operator has been proposed, which considers the similarity among the neighboring pixels and suggests a new texture coding scheme for texture classification. Experiments have been conducted on Brodatz and Outex database images and their performances have been noted. The proposed orthogonal based feature extraction scheme has been applied for fingerprint recognition and the performances have been reported. newline newline newline