Title

High order volumetric directional pattern for robust face recognition

Date of Award

2017

Degree Name

Ph.D. in Electrical and Computer Engineering

Department

Department of Electrical and Computer Engineering

Advisor/Chair

Advisor: Vijayan K. Asari

Abstract

The texture of objects in digital images is an important property that has been utilized in many computer vision and image analysis applications, such as pattern recognition, object classification, and region segmentation. Despite its frequent usage and many attempts to describe it in general terms, the texture lacks a precise definition. This makes the development of new texture descriptors a big challenge. In addition, researchers interest has recently spread into the dynamic texture (video domain), where the problem becomes more challenging. The main goal of feature description and representation techniques is to extract features from the image that are distinct and stable under different conditions during the image acquisition process. Texture descriptors can be generally classified into structural and statistical approaches. The structural methods consider the texture as a repetition of some primitives, with a specific rule of placement, while the statistical techniques characterize the stochastic properties of the spatial distribution of gray levels in an image using the gray tone co-occurrence matrix. In this work, we propose a combination of the structural and statistical approaches that can be utilized to recognize a variety of different textures, named High Order Local Directional Pattern (HOLDP) for still image based feature extraction (static texture) as well as High Order Volumetric Directional Pattern (HOVDP) for video based feature extraction (dynamic texture). Recently, the conventional Local Directional Pattern (LDP) has received a great deal of attention in face recognition applications. However, it only describes the micro structures of the texture images because it considers only a small neighborhood size. In fact, our proposed HOLDP descriptor can capture more detailed discriminative information by not only extracting the micro structures but also the macro structures of the texture images, which can be done by the help of a pyramidal multi-structure approach. The pyramid based multi-structure presented in this dissertation research can be created by encoding the directional information from different neighborhood layers of the image for each pixel position, and then concatenating the feature vectors of each neighborhood layer to form the final HOLDP feature map. Identifying human faces in video is a challenging problem due to the presence of large variations in facial pose and expression, as well as poor video resolution. To address this, Volumetric Directional Pattern (VDP) is proposed [1]. VDP is an oriented volumetric descriptor that is able to extract and fuse the information of multiple frames, temporal (dynamic) information, and multiple poses and expressions of faces in input videos to produce strong feature vectors. Meanwhile, to demonstrate the generality and capability of the HOLDP method, we develop another novel video based feature extraction technique, namely High Order Volumetric Directional Pattern (HOVDP) as an extension of VDP. HOVDP combines the movement and appearance features together by considering the nth order directional variation patterns of all neighboring pixel layers from three consecutive frames. From extensive experiments on still image based and video based face recognition benchmarks, we demonstrate the excellent performance of our proposed techniques compared to the state-of-the-art approaches.

Keywords

Image analysis, Computer vision, Pattern recognition systems, Human face recognition (Computer science), Engineering, Electrical Engineering, Face recognition, high order local directional pattern, volumetric directional pattern, high order volumetric directional pattern

Rights Statement

Copyright 2017, author

Share

COinS