http://ivylab.kaist.ac.kr/image/demo/project_title.gif

image004

image005

Color Texture Feature for Robust Face Recognition

2011. 06. 01 ~

 

image007

1. Introduction

l  Color information can provide complementary information to face recognition (FR) using texture analysis.

l  Most of FR approaches using color texture has been limited to encoding the texture patterns of only color pixel variations derived from each individual spectral band image.

l  We propose a new face descriptor so-called Local Color Vector Binary Patterns (LCVBP) for a FR purpose.

ž   LCVBP consists of two discriminant patterns: color norm patterns and color angular patterns.

ž   In particular, color angular patterns allow for encoding the discriminating texture patterns that are derived from spatial interactions among different spectral band images.

2. Extraction of Local Color Vector Binary Patterns

 

Fig. 1.  Defining the color vector of a pixel location within local regions of three channels in YIQ color space.

 

 

Fig. 2.  Illustration of extracting the LCVBP from YIQ color image.

 

l  Step 1: Define color vector  at a center pixel location  of each local region obtained from a set of  color band images .

l  Step 2: Compute color norm value  and color angle values  (between i-th and j-th bands) of , respectively.

l  Step 3: Define uniform LBP operators  and  for a color norm pattern and color angular patterns at location .

           

If

(1)

Otherwise

 

ž   : Color norm values of color vector  that form a circular neighborhood

ž   : Indexing function that labels a particular index to each of uniform patterns

ž   : Total number of transitions at  from 0 to 1 or vice versa

ž    are defined in a similar way by using the Equation of

l  Step 4: Compute LBP histograms of each local region  for color norm and color angular patterns.

l  Step 5: Generate LBP histograms of  for color norm and color angular patterns by separately concatenating regional LBP histograms obtained from associated patterns.

3. Face Recognition with LCVBP Feature

l  Low-dimensional feature extraction technique is adopted to protect FR performance from being degraded caused by high dimensionality and redundant information

For   

(2)

 

ž   : Feature extractors formed with LBP histograms of training face images for color norm and color angular patterns

ž   : LBP histograms of a probe RGB face image for color norm pattern and color angular patterns

ž   : Low-dimensional features of  and

l  Feature-level information fusion technique is used to combine a total of  features (one for color norm and  for color angular patterns) for facilitating complementary effect

(3)

 

l  LCVBP feature of a gallery RGB face image is created in the same way as probe RGB face image.

l  A nearest neighbor (NN) classifier is used for FR.

4. Experiment

<Experimental setup>

l  For comparison, FR methods using only grayscale texture (i.e., grayscale LBP, grayscale Gabor) and FR method using color texture (i.e., color LBP) were used

l  RQCr (R from RGB, and Q and Cr from YIQ and YCbCr color spaces) color configuration was used for extracting the LCVBP feature.

(a)

(b)

(c)

Fig. 3.  Examples of used face images. (a) CMU-PIE DB (with illumination change). (b) XM2VTS DB (with illumination change). (C) Color FERET DB (facial pose change).

<Experimental results>

Fig. 4.  Results on illumination variation (left) and pose variation (right).

 

l  Under severe illumination variation

ž   A total of 2,492 illuminant face images of 201 subjects were collected from the CMU-PIE and XM2VTS DB.

ž   The LCVBP feature is more robust against illumination variation than the grayscale LBP, grayscale Gabor, and color LBP.

l  Under pose variation

ž   A total of 1,378 face images of 107 subjects were collected from the Color FERET.

ž   The LCVBP feature is effective for FR under pose variation

5. Conclusions

l  In the proposed LCVBP, color norm and color angular patterns are extracted from a color face image

l  The LCVBP is highly robust to severe variation in illumination and moderate facial pose changes

* Contact Person: Prof. Yong Man Ro (ymro@kaist.ac.kr)

http://ivylab.kaist.ac.kr/image/demo/project_publication.gif

1.     S. H. Lee, J. Y. Choi, Y. M. Ro, and K. N. Plataniotis, Local Color Vector Binary Patterns from Multichannel Face Images for Face Recognition, IEEE Transactions on Image Processing, 2012.

2.     J. Y. Choi, Y. M. Ro, and K. N. Plataniotis, Color Local Texture Features for Color Face Recognition, IEEE Transactions on Image Processing, 2012.

3.     S. H. Lee, H. Kim, K. N. Plataniotis, and Y. M. Ro, Using Color Texture Sparsity for Facial Expression Recognition, IEEE International Conference on Automatic Face and Gesture Recognition, 2013.

4.     S. H. Lee, H. Kim, Y. M. Ro, A Comparative Study of Color Texture Features for Face Analysis, IAPR The Fourth Computational Color Imaging Workshop, 2013.

5.     S. H. Lee, J. Y. Choi, K. N. Plataniotis, and Y. M. Ro, Local Color Vector Binary Patterns for Face Recognition, IEEE Conference on Image Processing, 2011.

 

http://ivylab.kaist.ac.kr/image/demo/project_demo.gif

l  Video1: Effectiveness of color texture under illumination change.

 

l  Video2: Effectiveness of color texture for small resolution faces.