Advanced Recognition System

Advanced Eigenface Face Recognition

SystemPeixi Xiong

Northwestern University

Background• Facial Recognition is widely used since its convenience.

Motivation

Problems:

• Low Recognition Rate

• Time Consuming

Motivation

• Database in Varying Pose• Lack of Illumination• Glasses Interferences

• Database in Larger scale

Advantage and Drawback of the PCA Face Recognition

• Advantage • Drawback

Drawback • Time Consuming• Varying Pose and Illumination • Frontal Views/Face Expression and Disguise

Brief Analysis of the Database and the Original System • Good Illumination • Varying Pose (Tilted Face )• NOT Only the Frontal Views

The Pre-Processing System

Build Face Part Detector/Crop the Face• Output the Face Boundary, Left eye Boundary, Right eye Boundary,

Mouth Boundary and Nose Boundary

• The relevant information is included in the documentation of Matlab, vision.CascadeObjectDetector System object.

The Pre-Processing System

Detect If It Is Frontal Face• Detect Box • Check if Two Eyes Boxes Exist• Location of Mouth & Nose

Details-Types of Side Faces

The Second Type• Get Vertical Center of Mouth and Nose • Calculate Absolute Distance of Two Middle Lines • ‘Relative’ Distance • Set Threshold

The Pre-Processing System

• Check Location of Left Eye&Right Eye (the boundary box)• Get Center points• Link • Check the Angle • imrotate() • Rebuild Face Part Detector&Recrop the Face

Detect If It Is Tilted Face

The Pre-Processing System

Illumination-Correction • • Low Frequency Component & High Frequency Component • DCT Normalization (Discrete Cosine Transform Normalization)•

K. P. Horn, Robot Vision. Cambridge, MA: MIT Press, 1986.

Details• Installation • Do Image Histogram Normalization • Create & Save Zigzag Map • Transform to Logarithm & Frequency Domains• Set Zero to the Relevant Coefficient• Inverse DCT • Post-Processing

Result

Part II

Preprocessing • Crop the face[1]

• Convert RGB to Grayscale

Edge Detection• Not Continuous

Fill• Fill the eyeglasses inner part

• Why ? • Some flaw on edge detection

Method -- Imfill• How to Compensate?

Method--dilate• Expand the mask

Possible problem • Nose• Mouth • Hair• Other flaws

• delete the parts

Compensate• Recover masked skin • Smooth

Recover Masked Skin• Get mean within 5*5 box• Special case:• Not only one pixel in box• Recursively find the other pixels’ mean

Find Masked Pixel

Get Mean Value

OOPS!

Recursion

Ignore the

Recursion

¿

Recover & Smooth• Gaussian filter

Question?• How to remove wrong mask

Skin Color Detection:

Lighting Compensation

Color Transformastion

Skin Color Detection Variance-based Segmentation

Component Grouping

Face Boundary Detection

Resize Graph

Comparing

Experimental Results Analysis

• ROC Curve

• CMC Curve

ROC Curve

ROC Curve of advanced system is closer to the broader than original system

CMC Curve

CMC Curve of advanced system is closer to broader than traditional system.

Main Work • Delete Side-face• Crop Face• Tiled-face Rotation• Illumination Correction• Glasses Removal • Properties Analysis

Q&A

Thank you!