upm, faculty of computer science & it, 2009 1 a robust automated attendance system using face...

27
UPM, Faculty of Computer Science & IT, 2009 1 A robust automated attendance system A robust automated attendance system using face recognition techniques using face recognition techniques PhD proposal; May 2009 Gawed Nagi Supervisor Dr. Rahmita Wirza Supervisory committee Dr. Muhamad Taufik Dr. Fatimah Khalid

Post on 20-Dec-2015

213 views

Category:

Documents


0 download

TRANSCRIPT

UPM, Faculty of Computer Science & IT, 2009 1

A robust automated attendance system using A robust automated attendance system using face recognition techniquesface recognition techniques

PhD proposal; May 2009

Gawed Nagi

Supervisor

Dr. Rahmita Wirza

Supervisory committee

Dr. Muhamad TaufikDr. Fatimah Khalid

UPM, Faculty of Computer Science & IT, 2009 2

Contents

Introduction BackgroundProblem StatementResearch objectives Basic Assumptions Research significanceImplication Scope

Related Work Methodology Work Timetable References

UPM, Faculty of Computer Science & IT, 2009 3

Background

Face recognition (FR) is an active field in computer vision and many other areas.

The need for robust and accurate face recognition systems for security and commercial purposes is increasing.

Under uncontrolled conditions, FR is still a challenging task Partial occlusions, in face images are still an important

problem for most face recognition systems Normalizing detected face images can improve FR system

robustness and reliability

UPM, Faculty of Computer Science & IT, 2009 4

Problem Statement

Up to date, attendance system has been taken manually which causing time waste, paper work, besides it is inaccurate. Face recognition technology can be utilized to build an automated attendance system that makes counting and identifying students much easier and convenient. Face occlusions, face scaling, and posture are still important problems in such systems.

UPM, Faculty of Computer Science & IT, 2009 5

Research objectives

To establish simulation of all possible facial occlusions and face posture orientation

To determine the state-of-the-art techniques dealing with facial occlusions, face pose estimation, and face scale variation.

To build and evaluate a novel face recognition model with robustness against Common facial occlusions (glasses, scarves, hijab, ..) Face pose estimation Face scale variation

UPM, Faculty of Computer Science & IT, 2009 6

Basic Assumptions

Head view angle of subjects is between +15 o and – 15o

Cameras have high resolution (preferred 2896 x 1944 pixels)

UPM, Faculty of Computer Science & IT, 2009 7

Research significance

Robustness is still a big concern for real face recognition systems that work on face images taken under uncontrolled conditions including facial occlusions, pose changes, and face scale variation

More research is still needed to investigate and overcome those issues

Normalizing face images and applying ICA + RFE-SVM can improve the robustness of FR system and increasing the recognition rate with face occlusions, face scaling ,and posture.

UPM, Faculty of Computer Science & IT, 2009 8

Implication

AAS technology can be widely applied in private and public sectors to automatically performing the attendance in different time intervals.

Partial facial occlusions model can be applied to cope with facial occlusion problem in surveillance and security systems.

The dataset with different facial partial occlusions including Hijab will be made publicly available

UPM, Faculty of Computer Science & IT, 2009 9

Scope

Neutral (normal) face expression Controlled illumination Frontal/ near- frontal view faces

UPM, Faculty of Computer Science & IT, 2009 10

Related Work 1/2

UPM, Faculty of Computer Science & IT, 2009 11

Related Work 2/2

UPM, Faculty of Computer Science & IT, 2009 12

METHODOLOGY

The common scenarios of system design

Sample sizeNumber of images is 15 and The total number of subjects is 270

Scenario 2 Scenario 1

UPM, Faculty of Computer Science & IT, 2009 13

Cont..

Preliminary Experiment images

One sample (1420 X 528) of preliminary experiment pictures

UPM, Faculty of Computer Science & IT, 2009 14

Detected faces with OpenCV code

Resolution: 1420 X 528 pixels

Resolution: 710 X 264 pixels

UPM, Faculty of Computer Science & IT, 2009 15

Preliminary Experiment Results 1

UPM, Faculty of Computer Science & IT, 2009 16

Preliminary Experiment Results 2

UPM, Faculty of Computer Science & IT, 2009 17

The AAS Overview

Classified as Classified as ““KnownKnown””=> => ““presentpresent””

UPM, Faculty of Computer Science & IT, 2009 18

Data Collection

gathering 100 scene images for 5 different groups (each group is between 15 to 20 students) in two class sessions with 2-week time intervals. The collected images are including a round 1800 faces for approximately 75 distinct students (male and female with glasses, scarves, and Hijab, ..).

AR database, which includes two occlusion types sunglasses and scarves, will be used to test and evaluate the proposed techniques.

UPM, Faculty of Computer Science & IT, 2009 19

Face Detection (FD)

FD phase has a major influence on the performance and reliability of any face recognition system

The accurate FD is, the higher better face recognition is. Adaboost learning algorithm [Voila & Jones] is proposed for

face detection due to its several techniques for effective computation of a large number of features under varying scale and location, which is important for real time performance.

Two kinds of errors may occur in face detection:False negative when face not detectedFalse positive when non-face is detected

UPM, Faculty of Computer Science & IT, 2009 20

The contribution

Geometry face normalization The detected face varies in view angle, brightness,

size, and etc Those features are independent of face features and

will affect the recognition Rate significantly To improve the system recognition rate and real time

efficiency as well, the research proposes an algorithm to normalize face orientation and scaling using fourier transform.

UPM, Faculty of Computer Science & IT, 2009 21

Cont..

Feature Extraction and ClassificationFaces form a unique class of objects Feature extraction plays a very important role in any

pattern recognition system This research proposes a novel approach to cope with

varying types of facial occlusions based on

ICA + RFE-SVM.

UPM, Faculty of Computer Science & IT, 2009 22

ICA + RFE-SVM approach

Support Vector Machines (SVM) and Independent Component Analysis (ICA) are two powerful and relatively recent techniques

SVMs are classifiers which have demonstrated high generalization capabilities. ICA is a feature extraction technique which can be considered a generalization of Principal Component Analysis (PCA).

This approach is going to use ICA as a projection and feature extraction method, then selecting and ranking the features by employing Recursive Feature Elimination (RFE-SVM) so

the occluded features will gain the least weight which makes them removed. finally, classification is performed by SVM.

UPM, Faculty of Computer Science & IT, 2009 23

Why ICA? Why RFE-SVM?

Why ICA?

Why RFE-SVM?? Expected Results

We expect that our model will achieve a good recognition rate with robustness in terms efficiency, speed, and memory and storage requirements

A comparison between the recognition rate of faces with and without normalization method will be conducted.

A comparison between our model performance and the state-of-the-art models such as

– face recognition under partial occlusion based on support vector machine with local Gaussian summation kernel

UPM, Faculty of Computer Science & IT, 2009 24

Performance Evaluation

False reject error and false accept error True Accept Rate (TAR) and True Reject Rate (TRR) The balanced error rate (BER). Its definition is:

Receiver Operating Characteristics (ROC)

UPM, Faculty of Computer Science & IT, 2009 25

Time Table

UPM, Faculty of Computer Science & IT, 2009 26

References

UPM, Faculty of Computer Science & IT, 2009 27

Thank you for your attention