You are here

High Performance Techniques for Face Recognition

Download pdf | Full Screen View

Date Issued:
2017
Abstract/Description:
The identification of individuals using face recognition techniques is a challenging task. This is due to the variations resulting from facial expressions, makeup, rotations, illuminations, gestures, etc. Also, facial images contain a great deal of redundant information, which negatively affects the performance of the recognition system. The dimensionality and the redundancy of the facial features have a direct effect on the face recognition accuracy. Not all the features in the feature vector space are useful. For example, non-discriminating features in the feature vector space not only degrade the recognition accuracy but also increase the computational complexity.In the field of computer vision, pattern recognition, and image processing, face recognition has become a popular research topic. This is due to its wide spread applications in security and control, which allow the identified individual to access secure areas, personal information, etc. The performance of any recognition system depends on three factors: 1) the storage requirements, 2) the computational complexity, and 3) the recognition rates.Two different recognition system families are presented and developed in this dissertation. Each family consists of several face recognition systems. Each system contains three main steps, namely, preprocessing, feature extraction, and classification. Several preprocessing steps, such as cropping, facial detection, dividing the facial image into sub-images, etc. are applied to the facial images. This reduces the effect of the irrelevant information (background) and improves the system performance. In this dissertation, either a Neural Network (NN) based classifier or Euclidean distance is used for classification purposes. Five widely used databases, namely, ORL, YALE, FERET, FEI, and LFW, each containing different facial variations, such as light condition, rotations, facial expressions, facial details, etc., are used to evaluate the proposed systems. The experimental results of the proposed systems are analyzed using K-folds Cross Validation (CV).In the family-1, Several systems are proposed for face recognition. Each system employs different integrated tools in the feature extraction step. These tools, Two Dimensional Discrete Multiwavelet Transform (2D DMWT), 2D Radon Transform (2D RT), 2D or 3D DWT, and Fast Independent Component Analysis (FastICA), are applied to the processed facial images to reduce the dimensionality and to obtain discriminating features. Each proposed system produces a unique representation, and achieves less storage requirements and better performance than the existing methods.For further facial compression, there are three face recognition systems in the second family. Each system uses different integrated tools to obtain better facial representation. The integrated tools, Vector Quantization (VQ), Discrete cosine Transform (DCT), and 2D DWT, are applied to the facial images for further facial compression and better facial representation. In the systems using the tools VQ/2D DCT and VQ/ 2D DWT, each pose in the databases is represented by one centroid with 4*4*16 dimensions. In the third system, VQ/ Facial Part Detection (FPD), each person in the databases is represented by four centroids with 4*Centroids (4*4*16) dimensions. The systems in the family-2 are proposed to further reduce the dimensions of the data compared to the systems in the family-1 while attaining comparable results. For example, in family-1, the integrated tools, FastICA/ 2D DMWT, applied to different combinations of sub-images in the FERET database with K-fold=5 (9 different poses used in the training mode), reduce the dimensions of the database by 97.22% and achieve 99% accuracy. In contrast, the integrated tools, VQ/ FPD, in the family-2 reduce the dimensions of the data by 99.31% and achieve 97.98% accuracy. In this example, the integrated tools, VQ/ FPD, accomplished further data compression and less accuracy compared to those reported by FastICA/ 2D DMWT tools. Various experiments and simulations using MATLAB are applied. The experimental results of both families confirm the improvements in the storage requirements, as well as the recognition rates as compared to some recently reported methods.
Title: High Performance Techniques for Face Recognition.
22 views
7 downloads
Name(s): Aldhahab, Ahmed, Author
Mikhael, Wasfy, Committee Chair
Atia, George, Committee CoChair
Jones, W Linwood, Committee Member
Wei, Lei, Committee Member
Elshennawy, Ahmad, Committee Member
University of Central Florida, Degree Grantor
Type of Resource: text
Date Issued: 2017
Publisher: University of Central Florida
Language(s): English
Abstract/Description: The identification of individuals using face recognition techniques is a challenging task. This is due to the variations resulting from facial expressions, makeup, rotations, illuminations, gestures, etc. Also, facial images contain a great deal of redundant information, which negatively affects the performance of the recognition system. The dimensionality and the redundancy of the facial features have a direct effect on the face recognition accuracy. Not all the features in the feature vector space are useful. For example, non-discriminating features in the feature vector space not only degrade the recognition accuracy but also increase the computational complexity.In the field of computer vision, pattern recognition, and image processing, face recognition has become a popular research topic. This is due to its wide spread applications in security and control, which allow the identified individual to access secure areas, personal information, etc. The performance of any recognition system depends on three factors: 1) the storage requirements, 2) the computational complexity, and 3) the recognition rates.Two different recognition system families are presented and developed in this dissertation. Each family consists of several face recognition systems. Each system contains three main steps, namely, preprocessing, feature extraction, and classification. Several preprocessing steps, such as cropping, facial detection, dividing the facial image into sub-images, etc. are applied to the facial images. This reduces the effect of the irrelevant information (background) and improves the system performance. In this dissertation, either a Neural Network (NN) based classifier or Euclidean distance is used for classification purposes. Five widely used databases, namely, ORL, YALE, FERET, FEI, and LFW, each containing different facial variations, such as light condition, rotations, facial expressions, facial details, etc., are used to evaluate the proposed systems. The experimental results of the proposed systems are analyzed using K-folds Cross Validation (CV).In the family-1, Several systems are proposed for face recognition. Each system employs different integrated tools in the feature extraction step. These tools, Two Dimensional Discrete Multiwavelet Transform (2D DMWT), 2D Radon Transform (2D RT), 2D or 3D DWT, and Fast Independent Component Analysis (FastICA), are applied to the processed facial images to reduce the dimensionality and to obtain discriminating features. Each proposed system produces a unique representation, and achieves less storage requirements and better performance than the existing methods.For further facial compression, there are three face recognition systems in the second family. Each system uses different integrated tools to obtain better facial representation. The integrated tools, Vector Quantization (VQ), Discrete cosine Transform (DCT), and 2D DWT, are applied to the facial images for further facial compression and better facial representation. In the systems using the tools VQ/2D DCT and VQ/ 2D DWT, each pose in the databases is represented by one centroid with 4*4*16 dimensions. In the third system, VQ/ Facial Part Detection (FPD), each person in the databases is represented by four centroids with 4*Centroids (4*4*16) dimensions. The systems in the family-2 are proposed to further reduce the dimensions of the data compared to the systems in the family-1 while attaining comparable results. For example, in family-1, the integrated tools, FastICA/ 2D DMWT, applied to different combinations of sub-images in the FERET database with K-fold=5 (9 different poses used in the training mode), reduce the dimensions of the database by 97.22% and achieve 99% accuracy. In contrast, the integrated tools, VQ/ FPD, in the family-2 reduce the dimensions of the data by 99.31% and achieve 97.98% accuracy. In this example, the integrated tools, VQ/ FPD, accomplished further data compression and less accuracy compared to those reported by FastICA/ 2D DMWT tools. Various experiments and simulations using MATLAB are applied. The experimental results of both families confirm the improvements in the storage requirements, as well as the recognition rates as compared to some recently reported methods.
Identifier: CFE0006709 (IID), ucf:51878 (fedora)
Note(s): 2017-08-01
Ph.D.
Engineering and Computer Science, Electrical Engineering and Computer Engineering
Doctoral
This record was generated from author submitted information.
Subject(s): Face Recognition -- Discrete Wavelet Transform -- Discrete Multiwavelet Transform -- Discrete Radon Transform -- Discrete Cosine Transform -- Independent Component Analysis -- and Neural Network
Persistent Link to This Record: http://purl.flvc.org/ucf/fd/CFE0006709
Restrictions on Access: public 2017-08-15
Host Institution: UCF

In Collections