The Impact of Image Quality on the Performanceof Face Recognition
Abhishek Dutta1, Raymond Veldhuis, Luuk Spreeuwers
University of Twente, Netherlands.
May 25, 2012
33rd WIC Symposium on Information Theory in the Benelux, May 24-25, 2012
1http://abhishekdutta.org
1 / 28
Introduction
What is the impact of image quality on the performance of acommercial face recognition system2?
Motion Blur (angle = 0)
length = 03 length = 17
Gaussian Noise (mean= 0)
var. = 0.007 var. = 0.3
Pose
Illumination
60 x 45 120 x 90
Resolution
2Cognitec FaceVACS SDK 8.4.0
2 / 28
Introduction
What is the impact of image quality on the performance of acommercial face recognition system2?
Motion Blur (angle = 0)
length = 03 length = 17
Gaussian Noise (mean= 0)
var. = 0.007 var. = 0.3
Pose
Illumination
60 x 45 120 x 90
Resolution
2Cognitec FaceVACS SDK 8.4.0
3 / 28
A Quick Question: Fill in the blankFace recognition is
I difficult
I very difficult
[MultiPIE]
face recognition performance depends on image quality of the pairof images participating in the comparison process.
4 / 28
A Quick Question: Fill in the blankFace recognition is
I difficult
I very difficult
[MultiPIE]
face recognition performance depends on image quality of the pairof images participating in the comparison process.
5 / 28
A Quick Question: Fill in the blankFace recognition is
I difficult
I very difficult
[MultiPIE]
face recognition performance depends on image quality of the pairof images participating in the comparison process.
6 / 28
Measure of Face Recognition Performance
7 / 28
Measure of Face Recognition Performance
8 / 28
Measure of Face Recognition Performance
9 / 28
Measure of Face Recognition Performance
10 / 28
Pose and Illumination
* : Illumination (flash) placed just above the respective camera
-
6
Test Illumination*
Ref
.Il
lum
ina
tio
n*
11 / 28
Pose and Illumination
gradual performance degradation as ref. pose moves away from test pose12 / 28
Pose and Illumination
gradual performance degradation as ref. pose moves away from test pose13 / 28
Pose and Illumination ...
I gradual reduction in recognition performanceas the reference set pose moves away fromthe pose in the test set
I role of illumination insignificant for same pose(and camera3) in the test and reference set
Test and reference images having similar pose and captured bysame camera can greatly improve recognition performance.
3in the MultiPIE data set, if we exactly match the pose, we are alsomatching all the imaging characteristics
14 / 28
Pose and Illumination ...
I gradual reduction in recognition performanceas the reference set pose moves away fromthe pose in the test set
I role of illumination insignificant for same pose(and camera3) in the test and reference set
Test and reference images having similar pose and captured bysame camera can greatly improve recognition performance.
3in the MultiPIE data set, if we exactly match the pose, we are alsomatching all the imaging characteristics
15 / 28
Resolution
19_0 05_1
0.5
0.6
0.7
0.840x30
60x45
80x60
100x75
120x90
160x120
200x150
640x480
40x30
60x45
80x60
100x75
120x90
160x120
200x150
640x480
Reference set resolution
Are
a U
nder
RO
C (
AU
C)
Test set res.
(dist. between
eyes in pixels)
40x30 (13)
60x45 (16)
80x60 (18)
100x75 (20)
120x90 (22)
160x120 (25)
200x150 (28)
640x480 (50)
Motion Blur (angle = 0)
length = 03 length = 17
Gaussian Noise (mean= 0)
var. = 0.007 var. = 0.3
Pose
Illumination
60 x 45 120 x 90
Resolution
Tes
t&
ref.
po
se
16 / 28
Resolution ...
19_0 05_1
0.5
0.6
0.7
0.840x30
60x45
80x60
100x75
120x90
160x120
200x150
640x480
40x30
60x45
80x60
100x75
120x90
160x120
200x150
640x480
Reference set resolution
Are
a U
nder
RO
C (
AU
C)
Test set res.
(dist. between
eyes in pixels)
40x30 (13)
60x45 (16)
80x60 (18)
100x75 (20)
120x90 (22)
160x120 (25)
200x150 (28)
640x480 (50)
Tes
t&
ref.
po
se
Analysis:I recognition performance improves with the resolution of the
test and reference set (as expected).I relative difference in pose between the test and reference
image determines the extent of influence of resolution.17 / 28
Noise
19_0 05_1
0.5
0.6
0.7
0.8
0 0.007 0.03 0.07 0.1 0.3 0 0.007 0.03 0.07 0.1 0.3
Ref. set Gaussian noise variance (mean = 0)
Are
a U
nder
RO
C (
AU
C) Test set Gaussian
noise variance
(mean = 0)
0
0.007
0.03
0.07
0.1
0.3
Motion Blur (angle = 0)
length = 03 length = 17
Gaussian Noise (mean= 0)
var. = 0.007 var. = 0.3
Pose
Illumination
60 x 45 120 x 90
Resolution
Tes
t&
ref.
po
se
18 / 28
Noise ...
19_0 05_1
0.5
0.6
0.7
0.8
0 0.007 0.03 0.07 0.1 0.3 0 0.007 0.03 0.07 0.1 0.3
Ref. set Gaussian noise variance (mean = 0)
Are
a U
nder
RO
C (
AU
C) Test set Gaussian
noise variance
(mean = 0)
0
0.007
0.03
0.07
0.1
0.3
Tes
t&
ref.
po
se
Analysis:
I recognition performance degrades with the noise in the testand reference set (as expected).
I relative difference in pose between the test and referenceimage determines the extent of influence of zero meanGaussian noise.
19 / 28
Blur
19_0 05_1
0.60
0.65
0.70
0.75
0.80
0.85
0 3 5 7 13 17 0 3 5 7 13 17
Ref. set motion blur length (angle = 0)
Are
a U
nder
RO
C (
AU
C) Test set motion
blur length
(angle = 0)
0
3
5
7
13
17
Motion Blur (angle = 0)
length = 03 length = 17
Gaussian Noise (mean= 0)
var. = 0.007 var. = 0.3
Pose
Illumination
60 x 45 120 x 90
Resolution
Tes
t&
ref.
po
se
20 / 28
Blur ...
19_0 05_1
0.60
0.65
0.70
0.75
0.80
0.85
0 3 5 7 13 17 0 3 5 7 13 17
Ref. set motion blur length (angle = 0)
Are
a U
nder
RO
C (
AU
C) Test set motion
blur length
(angle = 0)
0
3
5
7
13
17
Tes
t&
ref.
po
se
Analysis:
I recognition performance degrades with the blur in the testand reference set (as expected).
I relative difference in pose between the test and referenceimage determines the extent of influence of image blur.
21 / 28
Summary
Quality Difference in Area Under ROC
Pose ∼ 50%Resolution ∼ 35%Noise (Gaussian) ∼ 35%Blur (Motion) ∼ 20%Illumination ∼ 20%
I approximately matching non-frontal pose between test andreference images (and using same camera) can greatlyimprove recognition performance.
I it is the relative difference in pose between the test andreference image that determines the extent of influence thatother quality parameters like illumination, noise, motion blur,and resolution have on the face recognition performance
I face recognition performance depends on quality of image pair
22 / 28
Summary
Quality Difference in Area Under ROC
Pose ∼ 50%Resolution ∼ 35%Noise (Gaussian) ∼ 35%Blur (Motion) ∼ 20%Illumination ∼ 20%
I approximately matching non-frontal pose between test andreference images (and using same camera) can greatlyimprove recognition performance.
I it is the relative difference in pose between the test andreference image that determines the extent of influence thatother quality parameters like illumination, noise, motion blur,and resolution have on the face recognition performance
I face recognition performance depends on quality of image pair
23 / 28
Summary
Quality Difference in Area Under ROC
Pose ∼ 50%Resolution ∼ 35%Noise (Gaussian) ∼ 35%Blur (Motion) ∼ 20%Illumination ∼ 20%
I approximately matching non-frontal pose between test andreference images (and using same camera) can greatlyimprove recognition performance.
I it is the relative difference in pose between the test andreference image that determines the extent of influence thatother quality parameters like illumination, noise, motion blur,and resolution have on the face recognition performance
I face recognition performance depends on quality of image pair
24 / 28
Limitations
I We have assumed that the image quality parameters areindependent. In reality, all the quality parameters co-exist andpresence or absence of one quality parameter (like pose, blur,etc) might affect the behavior of other quality parameters(like resolution, noise, etc)
I All the images used in this study were taken from a singleimage data set. Although test and reference images differedby session, ideally both test and reference images should havebeen taken from the different data set in order to simulate theconditions present in a real forensic case.
I these findings are limited by the inclusion of a specificcommercial face recognition system in this study.
25 / 28
Limitations
I We have assumed that the image quality parameters areindependent. In reality, all the quality parameters co-exist andpresence or absence of one quality parameter (like pose, blur,etc) might affect the behavior of other quality parameters(like resolution, noise, etc)
I All the images used in this study were taken from a singleimage data set. Although test and reference images differedby session, ideally both test and reference images should havebeen taken from the different data set in order to simulate theconditions present in a real forensic case.
I these findings are limited by the inclusion of a specificcommercial face recognition system in this study.
26 / 28
Limitations
I We have assumed that the image quality parameters areindependent. In reality, all the quality parameters co-exist andpresence or absence of one quality parameter (like pose, blur,etc) might affect the behavior of other quality parameters(like resolution, noise, etc)
I All the images used in this study were taken from a singleimage data set. Although test and reference images differedby session, ideally both test and reference images should havebeen taken from the different data set in order to simulate theconditions present in a real forensic case.
I these findings are limited by the inclusion of a specificcommercial face recognition system in this study.
27 / 28
Thank You
For more details, refer to:
The Impact of Image Quality on the Performance of Face Recognition, A. Dutta, R. Veldhuis, L. Spreeuwers, 33rd
WIC Symposium on Information Theory in the Benelux, May 2012.
28 / 28
Notes for Q/A Session
Table: Image quality variations included in this study.
QualityCamera Flash Resolution Motion Blur Gaus. Noise
Testi , Refj Testi , Refj Testi , Refj Testi , Refj Testi , RefjPose and Illumina-tion
ci , cj ∈ C fi , fj ∈ F ri , rj = D0 0, 0 0, 0
Resolution 19 1, {∗} 18, {∗∗} ri , rj ∈ R 0, 0 0, 0
Gaussian Noise 19 1, {∗} 18, {∗∗} ri , rj = D0 0, 0 σ̄i , σ̄j ∈ Nσ̄
Motion Blur 19 1, {∗} 18, {∗∗} ri , rj = D0 li , lj ∈ Bl 0, 0
C = [19 1, 19 0, 04 1, 05 0, 05 1, 14 0, 13 0, 08 0, 08 1],F = [02, 04, 14, 05, 15, 06, 07, 16, 08, 09, 17, 10, 18, 12], R = [640× 480, · · · , 60× 45],Bl (length in pixels) = [1, 3, 5, 7, 13, 17, 21], Note: angle= 0Nσ̄ (variance) = [0.001, 0.007, 0.03, 0.07, 0.1, 0.2]. Note: mean = 0{∗} = {19 1, 05 1}, {∗∗} = {10, 07}, D0 = 640× 480
Table: Test and reference set specifications (source: MultiPIE)
Test Set (Probe) Reference Set (Gallery)
size (image count) 479 442person count 319 268session 01,03 02,04
expression neutral neutraleye annotation manual manual
29 / 28
Notes for Q/A Session ...
05_0 04_1 19_0
19_108_1
08_0 13_0 14_0 05_1
chair withhead rest
04 05 0607
08 09 10 1202
14 15 16 1718
camera
flash
Figure: Camera and flash location for all the images used in thisexperiment (source : MultiPIE data set)
30 / 28