18
Matthias Wimmer, Bernd Radig, Michael Beetz Chair for Image Understanding Computer Science Technische Universität München [email protected] Adaptive Skin Color Classification

Matthias Wimmer, Bernd Radig, Michael Beetz Chair for Image Understanding Computer Science Technische Universität München [email protected] Adaptive

Embed Size (px)

Citation preview

Matthias Wimmer, Bernd Radig, Michael Beetz

Chair for Image UnderstandingComputer Science

Technische Universität München

[email protected]

Adaptive Skin Color Classification

2005-12-18 2/10Technische Universität MünchenMatthias Wimmer

Motivation

Skin color detection supports… face model fitting

mimic recognition person identification gaze estimation fatigue detection (e.g. vehicle)

hand tracking gesture recognition action recognition supervising work

challenge our approach results outlook

2005-12-18 3/10Technische Universität MünchenMatthias Wimmer

Challenge Skin color depends on image conditions:

illumination: light source, light color, shadow, shading,… camera: type, settings,… visible person: ethnic group, tan,…

Skin color occupies a large area within color space

challenge our approach results outlook

2005-12-18 4/10Technische Universität MünchenMatthias Wimmer

Observations Skin color varies greatly between images. Skin color varies slightly within an image.

Basic idea learn image specific skin color characteristics parameterize a skin color classifier accordingly

skin color of image1 skin color of image2

challenge our approach results outlook

2005-12-18 5/10Technische Universität MünchenMatthias Wimmer

Our approach

Offline step: learn the skin color mask

specific for any face detector

Online steps: Step 1: detect the image specific skin color model

using the face detector using the skin color mask

Step 2: adapt a skin color classifier Step 3: calculate the skin color image

challenge our approach results outlook

2005-12-18 6/10Technische Universität MünchenMatthias Wimmer

Skin color model & skin color classifier detect the face extract the skin color pixels skin color model:

mean values: μr, μg, μbase

standard deviations: σr, σg, σbase

adaptive skin color classifier:skin := lowr ≤ r ≤ highr

lowg ≤ g ≤ highg lowbase ≤ base ≤ highbase

learn the boundslowr := μr – 2σr

highr := μr + 2σr . . . . . . . . .

challenge our approach results outlook

2005-12-18 7/10Technische Universität MünchenMatthias Wimmer

Results better results for

colored persons exact shape outline detection of facial parts:

eyes, lips, brows,…

correctly detected pixels: non-adaptive approach: 90.4% 74.8%

40.2% adaptive approach: 97.5% 87.5%

97.0% improvement: 7.9% 17.0%

141.3%

challenge our approach results outlook

2005-12-18 8/10Technische Universität MünchenMatthias Wimmer

Conclusion Challenge: much variation within skin color

illumination, camera, visible person skin color occupies a large area within color space

We propose a way to reduce those variations exploit an image specific skin color model adapt a skin color classifier to that skin color model

We proved our approach using a simple but real-time capable skin color classifier comparison: non-adaptive ↔ adaptive

challenge our approach results outlook

2005-12-18 9/10Technische Universität MünchenMatthias Wimmer

Ongoing research Learn skin color mask for other face detectors Specialize more powerful skin color classifiers Recognize other feature images/color images

lip color image tooth color image eye color image hair color image eye brow color image

example: lip color detection

challenge our approach results outlook

2005-12-18 10/10Technische Universität MünchenMatthias Wimmer

Thank you !

2005-12-18 11/10Technische Universität MünchenMatthias Wimmer

Overview Motivation / Challenge Our approach

extracting skin color pixels adapt skin color classifier

Results Conclusion Outlook

2005-12-18 12/10Technische Universität MünchenMatthias Wimmer

Challenge (2): non-skin color pixels Skin color pixels have to be separated from non-

skin color pixels. Areas of skin color and

non-skin color overlap. Color can not make a

distinctive separation.

challenge our approach results outlook

2005-12-18 13/10Technische Universität MünchenMatthias Wimmer

Basic idea learn image specific skin color characteristics parameterize classifier accordingly with those

characteristics

2005-12-18 14/10Technische Universität MünchenMatthias Wimmer

Offline: Learn the skin color mask face image database with labeled skin color pixels skin color mask: array with 24 x 24 cells

Computational steps:

1. detect the face in every image

2. every cell is assigned the relative number of labeled skin color pixels at its position

3. apply threshold

1. 2. 3.

challenge our approach results outlook

2005-12-18 15/10Technische Universität MünchenMatthias Wimmer

Step 1: Detect the image specific skin color model

detect the face extract the skin color pixels normalized RGB color space:

base = R + G + B

r = R / base

g = G / base

skin color model: mean values: μr, μg, μbase

standard deviations: σr, σg, σbase

challenge our approach results outlook

2005-12-18 16/10Technische Universität MünchenMatthias Wimmer

Step 2: Adapt a skin color classifier non-adaptive skin color classifier:

skin := 0.35 ≤ r ≤ 0.5 0.2 ≤ g ≤ 0.7 200 ≤ base ≤ 740

adaptive skin color classifier:skin := lowr ≤ r ≤ highr

lowg ≤ g ≤ highg lowbase ≤ base ≤ highbase

learn the bounds via the skin color model mean value and standard deviationlowr := μr – 2σr

highr := μr + 2σr . . . . . . . . .

linear function:lowr := aμr + bμg + cμbase + dσr + eσg + fσbase + g

. . .

challenge our approach results outlook

2005-12-18 17/10Technische Universität MünchenMatthias Wimmer

Adaptive skin color classifier non adaptive skin color classifier:

skin := 0.35 ≤ r ≤ 0.5 0.2 ≤ g ≤ 0.7 200 ≤ base ≤ 740

adaptive skin color classifier:skin := lowr ≤ r ≤ highr lowg ≤ g ≤ highg

lowbase ≤ base ≤ highbase

learn the bounds out of the skin color model

2005-12-18 18/10Technische Universität MünchenMatthias Wimmer

Related work Feedback of information from

high level vision components to low level vision components

challenge our approach results outlook