Automatic Iris Segmentation Using Active Near Infra Red Lighting - - PowerPoint PPT Presentation

Automatic Iris Segmentation Using Active Near Infra Red Lighting

Carlos Morimoto Thiago Santos Adriano Saturno

Laboratory of Technologies for Interaction (LaTIn) Institute of Mathematics and Statistics (IME) University of S˜ ao Paulo (USP) hitoshi,thsant,saturno@ime.usp.br

SIBGRAPI 2005, Natal - RN - Brazil

www.ime.usp.br/∼thsant/sibgrapi05 (IME) Iris Segmentation using NIR SIBGRAPI’05 1 / 22

Teaser

We present here a technique for iris segmentation that uses infra-red light in the image acquisition step. The pupil is identified and regions of iris occluded by eyelids are removed. The system can work in low cost adapted hardware.

www.ime.usp.br/∼thsant/sibgrapi05 (IME) Iris Segmentation using NIR SIBGRAPI’05 2 / 22

Outline

1

Introduction

2

Acquisition with NIR light

3

Segmentation: pupil, iris and eyelids

4

Implementation and results

5

Conclusion

www.ime.usp.br/∼thsant/sibgrapi05 (IME) Iris Segmentation using NIR SIBGRAPI’05 3 / 22

Introduction

Iris as an identification method

Unique pattern per person

Twins don’t present the same pattern Phenotypical feature Two identical irises: 1 chance in 1078

www.ime.usp.br/∼thsant/sibgrapi05 (IME) Iris Segmentation using NIR SIBGRAPI’05 4 / 22

Introduction

Iris as an identification method

Unique pattern per person

Twins don’t present the same pattern Phenotypical feature Two identical irises: 1 chance in 1078

Stable along life

Structure: 3rd

¯ to 8th ¯ month of gestation

Pigment changes: until the 1st

¯ year of life www.ime.usp.br/∼thsant/sibgrapi05 (IME) Iris Segmentation using NIR SIBGRAPI’05 4 / 22

Introduction

Iris as an identification method

Unique pattern per person

Twins don’t present the same pattern Phenotypical feature Two identical irises: 1 chance in 1078

Stable along life

Structure: 3rd

¯ to 8th ¯ month of gestation

Pigment changes: until the 1st

¯ year of life

Iris Recognition Technology (IRT)

Pilot ATMs in USA and England Authentication in European airports Commercial systems available:

Evermedia LG Panasonic...

www.ime.usp.br/∼thsant/sibgrapi05 (IME) Iris Segmentation using NIR SIBGRAPI’05 4 / 22

Introduction

Iris as an identification method

Unique pattern per person

Twins don’t present the same pattern Phenotypical feature Two identical irises: 1 chance in 1078

Stable along life

Structure: 3rd

¯ to 8th ¯ month of gestation

Pigment changes: until the 1st

¯ year of life

Iris Recognition Technology (IRT)

Pilot ATMs in USA and England Authentication in European airports Commercial systems available:

Evermedia LG Panasonic...

IRT considered mature for biometry

www.ime.usp.br/∼thsant/sibgrapi05 (IME) Iris Segmentation using NIR SIBGRAPI’05 4 / 22

Introduction

Problems with IRT adoption

Training to use the system

positioning calibration

www.ime.usp.br/∼thsant/sibgrapi05 (IME) Iris Segmentation using NIR SIBGRAPI’05 5 / 22

Introduction

Problems with IRT adoption

Training to use the system

positioning calibration

Difficult to accomplish using low-cost equipment:

fixed focus micro cameras web-cams

www.ime.usp.br/∼thsant/sibgrapi05 (IME) Iris Segmentation using NIR SIBGRAPI’05 5 / 22

Introduction

Problems with IRT adoption

Training to use the system

positioning calibration

Difficult to accomplish using low-cost equipment:

fixed focus micro cameras web-cams

Acquisition problems:

eyelids occlusion specular reflection images out of focus

www.ime.usp.br/∼thsant/sibgrapi05 (IME) Iris Segmentation using NIR SIBGRAPI’05 5 / 22

Introduction

Problems with IRT adoption

Training to use the system

positioning calibration

Difficult to accomplish using low-cost equipment:

fixed focus micro cameras web-cams

Acquisition problems:

eyelids occlusion specular reflection images out of focus

Need a robust, efficient, easy to use and low-cost system for iris segmentation.

www.ime.usp.br/∼thsant/sibgrapi05 (IME) Iris Segmentation using NIR SIBGRAPI’05 5 / 22

Introduction

Simple model for iris

We would like to segment iris identifying

the pupil as a circle, the iris as a circle

• r its outer border - the limbus - and

the eyelids as elliptical segments.

www.ime.usp.br/∼thsant/sibgrapi05 (IME) Iris Segmentation using NIR SIBGRAPI’05 6 / 22

Introduction

Simple model for iris

We would like to segment iris identifying

the pupil as a circle, the iris as a circle

• r its outer border - the limbus - and

the eyelids as elliptical segments.

The eyelids position is arbitrary

www.ime.usp.br/∼thsant/sibgrapi05 (IME) Iris Segmentation using NIR SIBGRAPI’05 6 / 22

Acquisition

Camera

NIR light sources position

L1 and L2 turned on alternately

www.ime.usp.br/∼thsant/sibgrapi05 (IME) Iris Segmentation using NIR SIBGRAPI’05 7 / 22

Acquisition

Camera

NIR light sources position

L1 and L2 turned on alternately L1 → “red eye” effect: bright pupil

www.ime.usp.br/∼thsant/sibgrapi05 (IME) Iris Segmentation using NIR SIBGRAPI’05 7 / 22

Acquisition

System’s geometry

Near Infra Red light (NIR): 700-900nm.

Practically invisible to human eye

www.ime.usp.br/∼thsant/sibgrapi05 (IME) Iris Segmentation using NIR SIBGRAPI’05 8 / 22

Acquisition

System’s geometry

Near Infra Red light (NIR): 700-900nm.

Practically invisible to human eye

Light source L near to projection center C:

www.ime.usp.br/∼thsant/sibgrapi05 (IME) Iris Segmentation using NIR SIBGRAPI’05 8 / 22

Acquisition

System’s geometry

Near Infra Red light (NIR): 700-900nm.

Practically invisible to human eye

Light source L near to projection center C:

C, L and the reflection R (glint) are approximately collinears.

www.ime.usp.br/∼thsant/sibgrapi05 (IME) Iris Segmentation using NIR SIBGRAPI’05 8 / 22

Acquisition

System’s geometry

Near Infra Red light (NIR): 700-900nm.

Practically invisible to human eye

Light source L near to projection center C:

C, L and the reflection R (glint) are approximately collinears. R near the pupil center P if user is looking to the camera.

www.ime.usp.br/∼thsant/sibgrapi05 (IME) Iris Segmentation using NIR SIBGRAPI’05 8 / 22

Segmentation

System processing flow

www.ime.usp.br/∼thsant/sibgrapi05 (IME) Iris Segmentation using NIR SIBGRAPI’05 9 / 22

Segmentation Pupil

Outline

1

Introduction

2

Acquisition with NIR light

3

Segmentation: pupil, iris and eyelids Pupil segmentation Limbus and eyelids segmentation

4

Implementation and results

5

Conclusion

www.ime.usp.br/∼thsant/sibgrapi05 (IME) Iris Segmentation using NIR SIBGRAPI’05 10 / 22

Segmentation Pupil

Pupil segmentation

L1 produces a bright pupil image

Small brightness changes out pupil

www.ime.usp.br/∼thsant/sibgrapi05 (IME) Iris Segmentation using NIR SIBGRAPI’05 11 / 22

Segmentation Pupil

Pupil segmentation

L1 produces a bright pupil image

Small brightness changes out pupil

L2 produces an image with a dark pupil

www.ime.usp.br/∼thsant/sibgrapi05 (IME) Iris Segmentation using NIR SIBGRAPI’05 11 / 22

Segmentation Pupil

Pupil segmentation

L1 produces a bright pupil image

Small brightness changes out pupil

L2 produces an image with a dark pupil Images subtraction gets pupil’s pixels

www.ime.usp.br/∼thsant/sibgrapi05 (IME) Iris Segmentation using NIR SIBGRAPI’05 11 / 22

Segmentation Pupil

Pupil segmentation

L1 produces a bright pupil image

Small brightness changes out pupil

L2 produces an image with a dark pupil Images subtraction gets pupil’s pixels Circle fitting is used to segment the pupil

www.ime.usp.br/∼thsant/sibgrapi05 (IME) Iris Segmentation using NIR SIBGRAPI’05 11 / 22

Segmentation Pupil

Pupil segmentation

L1 produces a bright pupil image

Small brightness changes out pupil

L2 produces an image with a dark pupil Images subtraction gets pupil’s pixels Circle fitting is used to segment the pupil User looking to the camera → glint is

• ver the pupil

www.ime.usp.br/∼thsant/sibgrapi05 (IME) Iris Segmentation using NIR SIBGRAPI’05 11 / 22

Segmentation Pupil

Pupil segmentation

www.ime.usp.br/∼thsant/sibgrapi05 (IME) Iris Segmentation using NIR SIBGRAPI’05 11 / 22

Segmentation Limbus and eyelid

Outline

1

Introduction

2

Acquisition with NIR light

3

Segmentation: pupil, iris and eyelids Pupil segmentation Limbus and eyelids segmentation

4

Implementation and results

5

Conclusion

www.ime.usp.br/∼thsant/sibgrapi05 (IME) Iris Segmentation using NIR SIBGRAPI’05 12 / 22

Segmentation Limbus and eyelid

Limbus segmentation

Limbus orientation

Sclera is brighter than iris

Gradient taken at limbus points to sclera

www.ime.usp.br/∼thsant/sibgrapi05 (IME) Iris Segmentation using NIR SIBGRAPI’05 13 / 22

Segmentation Limbus and eyelid

Limbus segmentation

Limbus orientation

−1 1 −2 2 −1 1 Sclera is brighter than iris

Gradient taken at limbus points to sclera

Sobel operator used to get vertical edges

www.ime.usp.br/∼thsant/sibgrapi05 (IME) Iris Segmentation using NIR SIBGRAPI’05 13 / 22

Segmentation Limbus and eyelid

Limbus segmentation

Limbus orientation

−1 1 −2 2 −1 1 Sclera is brighter than iris

Gradient taken at limbus points to sclera

Sobel operator used to get vertical edges

Horizontal borders are not considered because eyelids occlusion

www.ime.usp.br/∼thsant/sibgrapi05 (IME) Iris Segmentation using NIR SIBGRAPI’05 13 / 22

Segmentation Limbus and eyelid

Limbus segmentation

Limbus orientation

−1 1 −2 2 −1 1 Sclera is brighter than iris

Gradient taken at limbus points to sclera

Sobel operator used to get vertical edges

Horizontal borders are not considered because eyelids occlusion

Referring to pupil center:

www.ime.usp.br/∼thsant/sibgrapi05 (IME) Iris Segmentation using NIR SIBGRAPI’05 13 / 22

Segmentation Limbus and eyelid

Limbus segmentation

Limbus orientation

−1 1 −2 2 −1 1 Sclera is brighter than iris

Gradient taken at limbus points to sclera

Sobel operator used to get vertical edges

Horizontal borders are not considered because eyelids occlusion

Referring to pupil center:

Left → negative edges

www.ime.usp.br/∼thsant/sibgrapi05 (IME) Iris Segmentation using NIR SIBGRAPI’05 13 / 22

Segmentation Limbus and eyelid

Limbus segmentation

Limbus orientation

−1 1 −2 2 −1 1 Sclera is brighter than iris

Gradient taken at limbus points to sclera

Sobel operator used to get vertical edges

Horizontal borders are not considered because eyelids occlusion

Referring to pupil center:

Left → negative edges Right → positive edges

www.ime.usp.br/∼thsant/sibgrapi05 (IME) Iris Segmentation using NIR SIBGRAPI’05 13 / 22

Segmentation Limbus and eyelid

Limbus segmentation

Limbus orientation

www.ime.usp.br/∼thsant/sibgrapi05 (IME) Iris Segmentation using NIR SIBGRAPI’05 13 / 22

Segmentation Limbus and eyelid

Limbus segmentation

Circle fitting

Canny operator used to get narrow and continuous edges

www.ime.usp.br/∼thsant/sibgrapi05 (IME) Iris Segmentation using NIR SIBGRAPI’05 14 / 22

Segmentation Limbus and eyelid

Limbus segmentation

Circle fitting

Canny operator used to get narrow and continuous edges A region around the pupil is fixed

www.ime.usp.br/∼thsant/sibgrapi05 (IME) Iris Segmentation using NIR SIBGRAPI’05 14 / 22

Segmentation Limbus and eyelid

Limbus segmentation

Circle fitting

Canny operator used to get narrow and continuous edges A region around the pupil is fixed The orientation constraints (previous step) get a set of points

www.ime.usp.br/∼thsant/sibgrapi05 (IME) Iris Segmentation using NIR SIBGRAPI’05 14 / 22

Segmentation Limbus and eyelid

Limbus segmentation

Circle fitting

www.ime.usp.br/∼thsant/sibgrapi05 (IME) Iris Segmentation using NIR SIBGRAPI’05 14 / 22

Segmentation Limbus and eyelid

Limbus segmentation

Circle fitting

Canny operator used to get narrow and continuous edges A region around the pupil is fixed The orientation constraints (previous step) get a set of points Minimum squares used to approximate a circumference

Hough transform and deformable models would be other options

www.ime.usp.br/∼thsant/sibgrapi05 (IME) Iris Segmentation using NIR SIBGRAPI’05 14 / 22

Segmentation Limbus and eyelid

Limbus segmentation

Circle fitting

www.ime.usp.br/∼thsant/sibgrapi05 (IME) Iris Segmentation using NIR SIBGRAPI’05 14 / 22

Segmentation Limbus and eyelid

Limbus segmentation

Successive pyramidal approximation

Limbus’ edges are more salient in low resolutions

www.ime.usp.br/∼thsant/sibgrapi05 (IME) Iris Segmentation using NIR SIBGRAPI’05 15 / 22

Segmentation Limbus and eyelid

Limbus segmentation

Successive pyramidal approximation

www.ime.usp.br/∼thsant/sibgrapi05 (IME) Iris Segmentation using NIR SIBGRAPI’05 15 / 22

Segmentation Limbus and eyelid

Limbus segmentation

Successive pyramidal approximation

pyramid up

Limbus’ edges are more salient in low resolutions Gaussian pyramid used to get multi-resolution images

www.ime.usp.br/∼thsant/sibgrapi05 (IME) Iris Segmentation using NIR SIBGRAPI’05 15 / 22

Segmentation Limbus and eyelid

Limbus segmentation

Successive pyramidal approximation

pyramid up

www.ime.usp.br/∼thsant/sibgrapi05 (IME) Iris Segmentation using NIR SIBGRAPI’05 15 / 22

Segmentation Limbus and eyelid

Limbus segmentation

Successive pyramidal approximation

pyramid up

Limbus’ edges are more salient in low resolutions Gaussian pyramid used to get multi-resolution images Results from low resolutions used to improve high resolution ones

www.ime.usp.br/∼thsant/sibgrapi05 (IME) Iris Segmentation using NIR SIBGRAPI’05 15 / 22

Segmentation Limbus and eyelid

Limbus segmentation

Successive pyramidal approximation

pyramid up

Limbus’ edges are more salient in low resolutions Gaussian pyramid used to get multi-resolution images Results from low resolutions used to improve high resolution ones Low resolution result used as mask

pyramid up source logical and circle fitting www.ime.usp.br/∼thsant/sibgrapi05 (IME) Iris Segmentation using NIR SIBGRAPI’05 15 / 22

Segmentation Limbus and eyelid

Limbus segmentation

Successive pyramidal approximation

pyramid up source logical and circle fitting

www.ime.usp.br/∼thsant/sibgrapi05 (IME) Iris Segmentation using NIR SIBGRAPI’05 15 / 22

Segmentation Limbus and eyelid

Eyelids segmentation

Similar strategy is used to eyelids.

www.ime.usp.br/∼thsant/sibgrapi05 (IME) Iris Segmentation using NIR SIBGRAPI’05 16 / 22

Segmentation Limbus and eyelid

Eyelids segmentation

−1 −2 −1 1 2 1 Similar strategy is used to eyelids. Sobel operator gets horizontal edges

www.ime.usp.br/∼thsant/sibgrapi05 (IME) Iris Segmentation using NIR SIBGRAPI’05 16 / 22

Segmentation Limbus and eyelid

Eyelids segmentation

−1 −2 −1 1 2 1 Similar strategy is used to eyelids. Sobel operator gets horizontal edges Canny operator, eyelid orientation and the limbus position (previously detected) combined to get candidate points

Column by columns, the nearest to pupil center points are picked

www.ime.usp.br/∼thsant/sibgrapi05 (IME) Iris Segmentation using NIR SIBGRAPI’05 16 / 22

Segmentation Limbus and eyelid

Eyelids segmentation

−1 −2 −1 1 2 1 Similar strategy is used to eyelids. Sobel operator gets horizontal edges Canny operator, eyelid orientation and the limbus position (previously detected) combined to get candidate points

Column by columns, the nearest to pupil center points are picked

Ellipse fitting applied

www.ime.usp.br/∼thsant/sibgrapi05 (IME) Iris Segmentation using NIR SIBGRAPI’05 16 / 22

Results

Results

Preliminary results gotten by the prototype

www.ime.usp.br/∼thsant/sibgrapi05 (IME) Iris Segmentation using NIR SIBGRAPI’05 17 / 22

Results

Requirements to get good images

Desired constraints imposed by the presented system

Real pupil: pictures and photographs don’t produces bright pupils (“red eye” effect).

www.ime.usp.br/∼thsant/sibgrapi05 (IME) Iris Segmentation using NIR SIBGRAPI’05 18 / 22

Results

Requirements to get good images

Desired constraints imposed by the presented system

Real pupil: pictures and photographs don’t produces bright pupils (“red eye” effect). Glint over the pupil: grants the user is looking to the camera.

www.ime.usp.br/∼thsant/sibgrapi05 (IME) Iris Segmentation using NIR SIBGRAPI’05 18 / 22

Results

Requirements to get good images

Desired constraints imposed by the presented system

Real pupil: pictures and photographs don’t produces bright pupils (“red eye” effect). Glint over the pupil: grants the user is looking to the camera. Pupil center near image center: reduces camera and perspective distortions.

www.ime.usp.br/∼thsant/sibgrapi05 (IME) Iris Segmentation using NIR SIBGRAPI’05 18 / 22

Results

Requirements to get good images

Desired constraints imposed by the presented system

Real pupil: pictures and photographs don’t produces bright pupils (“red eye” effect). Glint over the pupil: grants the user is looking to the camera. Pupil center near image center: reduces camera and perspective distortions. Corneal reflection with narrow area: grants input image is on focus.

www.ime.usp.br/∼thsant/sibgrapi05 (IME) Iris Segmentation using NIR SIBGRAPI’05 18 / 22

Results

Is the user looking to the camera?

Glint as a clue

www.ime.usp.br/∼thsant/sibgrapi05 (IME) Iris Segmentation using NIR SIBGRAPI’05 19 / 22

Conclusion

Conclusions and further work

The system is robust to environment illumination changes

www.ime.usp.br/∼thsant/sibgrapi05 (IME) Iris Segmentation using NIR SIBGRAPI’05 20 / 22

Conclusion

Conclusions and further work

The system is robust to environment illumination changes Automatically chooses better input images for segmentation:

positioning direction focus

www.ime.usp.br/∼thsant/sibgrapi05 (IME) Iris Segmentation using NIR SIBGRAPI’05 20 / 22

Conclusion

Conclusions and further work

The system is robust to environment illumination changes Automatically chooses better input images for segmentation:

positioning direction focus

Eyelid detection not so accurated yet

www.ime.usp.br/∼thsant/sibgrapi05 (IME) Iris Segmentation using NIR SIBGRAPI’05 20 / 22

Conclusion

Conclusions and further work

The system is robust to environment illumination changes Automatically chooses better input images for segmentation:

positioning direction focus

Eyelid detection not so accurated yet Further work:

Detection and elimination of eyelashes Interface to help user in the acquisition step

www.ime.usp.br/∼thsant/sibgrapi05 (IME) Iris Segmentation using NIR SIBGRAPI’05 20 / 22

Conclusion

References

• J. Daugman

High confidence visual recognition of persons by a test of statistical dependences IEEE Trans. on Pattern Analysis and Machine Inteligencem 15(11):1148–1161, 1993

• R. Wildes et al.

A system for automated iris recognition Applications of Computer Vision, Proceedings of the Second IEEE Workshop on (1994), pp. 121-128.

www.ime.usp.br/∼thsant/sibgrapi05 (IME) Iris Segmentation using NIR SIBGRAPI’05 21 / 22

Conclusion

Questions?

www.ime.usp.br/∼thsant/sibgrapi05 (IME) Iris Segmentation using NIR SIBGRAPI’05 22 / 22