smile detection by boosting pixel differences

22
Smile Detection by Boosting Pixel Differences Caifeng Shan, Member, IEEE IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 21, NO. 1, JANUARY 2012

Upload: kelvin

Post on 22-Feb-2016

41 views

Category:

Documents


0 download

DESCRIPTION

Smile Detection by Boosting Pixel Differences. Caifeng Shan , Member, IEEE IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 21, NO. 1, JANUARY 2012. Outline. INTRODUCTION METHOD EXPERIMENTS. Outline. INTRODUCTION METHOD EXPERIMENTS. INTRODUCTION. - PowerPoint PPT Presentation

TRANSCRIPT

Patch-Based Background Initialization in Heavily Cluttered Video

Smile Detection by Boosting Pixel DifferencesCaifeng Shan, Member, IEEE

IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 21, NO. 1, JANUARY 2012OutlineINTRODUCTION

METHOD

EXPERIMENTS

OutlineINTRODUCTION

METHOD

EXPERIMENTS

INTRODUCTIONMost of the existing works have been focused on analyzing a set of prototypic emotional facial expressions

Using the data collected by asking subjects to pose deliberately these expressions

In this paper, we focus on smile detection in face images captured in real-world scenariosINTRODUCTION

OutlineINTRODUCTION

METHOD

EXPERIMENTS

METHODBOOSTING PIXEL DIFFERENCES

S. Baluja and H. A. Rowley, Boosting set identification performance,Int. J. Comput. Vis., vol. 71, no. 1, pp. 111119, 2007

Baluja introduced to use the relationship between two pixels intensities as features.METHODthey used five types of pixel comparison operators (and their inverses):

METHODThe binary result of each comparison, which is represented numerically as 1 or 0, is used as the feature. Thus, for an image of pixels, there are or 3312000 pixel-comparison features

METHODInstead of utilizing the above comparison operators, we propose to use the intensity difference between two pixels as a simple feature

For an image of 24*24 pixels, there are or 331200 features extracted

METHOD AdaBoost ( Adaptive Boosting )AdaBoost learns a small number of weak classifiers whose performance is just better than random guessing and boosts them iteratively into a strong classifier of higher accuracy

the weak classifier consists of feature (i.e., the intensity difference),threshold , and parity indicating the direction of the inequality sign as follows:

METHOD

METHOD

OutlineINTRODUCTION

METHOD

EXPERIMENTS

EXPERIMENTS Data

Database : GENKI4K consists of 4000 images (2162 smile and 1828 nonsmile)

In our experiments, the images were converted to grayscale

the faces were normalized to reach a canonical face of 48*48 pixels

EXPERIMENTS Data

EXPERIMENTSIllumination NormalizationHistogram equalization (HE)

Single-scale retinex (SSR)

Discrete cosine transform (DCT)

LBP

TanTriggs

EXPERIMENTSIllumination Normalization

EXPERIMENTS Boosting Pixel Intensity Differences

Average of (left) all smile faces and (right) all nonsmile faces

EXPERIMENTSImpact of Pose Variation

EXPERIMENTSImpact of Pose Variation

Thank you