Digital Watermarking With Phase Dispersion Algorithm Team 1 Final Presentation SIMG 786 Advanced Digital Image Processing Mahdi Nezamabadi, Chengmeng Liu,

Digital Watermarking With Digital Watermarking With Phase Dispersion AlgorithmPhase Dispersion Algorithm

Team 1 Final PresentationTeam 1 Final Presentation

SIMG 786 SIMG 786

Advanced Digital Image ProcessingAdvanced Digital Image Processing

Mahdi Nezamabadi, Mahdi Nezamabadi,

Chengmeng Liu, Chengmeng Liu,

Michael SuMichael Su

Motivating ScenarioMotivating ScenarioMotivating ScenarioMotivating Scenario

• Alice creates a 3D shape,and publishes it on the web.

• Bob sells it as his own.Bob sells it as his own.

• How can Alice prove ownership?How can Alice prove ownership?(and make Bob pay her a lot of (and make Bob pay her a lot of money)money)

• Alice creates a 3D shape,and publishes it on the web.

• Bob sells it as his own.Bob sells it as his own.

• How can Alice prove ownership?How can Alice prove ownership?(and make Bob pay her a lot of (and make Bob pay her a lot of money)money)

The solution is…The solution is…

• An invisible, robust digital watermark and put it on the image which can be used for proving the ownership.

• It has been applied in copyright marking business.

• It can be also applied for digital multimedia

Digital Watermarking With Phase Digital Watermarking With Phase Dispersion AlgorithmDispersion Algorithm

• An algorithm for robust, invisible watermarking.• Use the spread-spectrum technique which was

first in communications for hiding the information.

• Uses this characteristics to hide and extract information.

• It can embed both iconic images and binary strings in an image.

• It can handle various types of attacks.

Malicious AttacksMalicious AttacksMalicious AttacksMalicious Attacks

Adding noise

Adding another watermark

Rescale

Lossy compression

Geometric distortion

Cropping

Print and scan

Adding noise

Adding another watermark

Rescale

Lossy compression

Geometric distortion

Cropping

Print and scan

Embedding process illustrationEmbedding process illustration

Watermark extraction processWatermark extraction process

Indices for image differenceIndices for image difference

• MSE (Mean square error)

2

1 1

1

N

i

M

j

i,jIi,jI'NM

I,I'MSE

• Correlation factor

),(),(

),(),('

yxMyxM

yxMyxMCF

Similarity vs. α

• Similarity is measured by cross correlation between original and extracted log

• 64 tiles were used in embedding

• The α controls the visibility of the watermark logo in the watermarked image

• The α also depends on the number of tiles

Implementation of Binary Message Implementation of Binary Message template function 1template function 1

• embedding binary information consists of representing the one and zero bits by positive delta function and black that are placed in predefined and unique locations within the message image.

• It consisted of concentric circles with equal increments in radius and random angular displacement.

• A 64 bits template is shown on left

• The error rate is 0 for this 64 bits template

Implementation of Binary Message Implementation of Binary Message template function 2template function 2

• 650 bits template function is shown on the left

• 650 bits can embed 32 characters by repeating them 5 times with no compression

• The error rate is 0.46% for this 650 bits template, that means the probability for get a wrong bit is 9.7e-8

Rotation/Scale DetectionThresholding

Rotation/Scale DetectionImage rotation

Rotation/Scale DetectionImage rotation

Robustness to lossy compressionRobustness to lossy compression

Original size Resolution MSE Correlation Factor

4.1MB 2k X 2k pixels 0.1194 0.5130

Compressed size

Compression ratio

555KB 7 0.1385 0.4172

312KB 13 0.1562 0.3798

199KB 20 0.1901 0.3251

Attacked by low pass filter

The watermarked image is blurred The extracted logo is equivalent to original log convolve with a low pass filter

Robustness with noiseRobustness with noise

Multiple watermarksMultiple watermarks

With the same keyEmbedded and extracted with different keys

Robustness to CroppingRobustness to Cropping

Halftoning can destroy the Halftoning can destroy the correlation between image and correlation between image and

watermarkwatermarkLena after printed and scanned Extracted watermark

ConclusionsConclusions

• This algorithm works best under the following circumstances:

• α = 0.2 gives the best balance between visibility and signal strength.

• Bigger image size and smaller watermarks ( more tiles).• Bigger color depths.• The algorithm can resist the following attacks: lowpass

filtering, cropping, noise, lossy compression, rotation, rescaling.

• But it does not handle halftoning.• It is sensitive to rotation angles.

Future workFuture work

• Deal with printer halftoning attacks

• Support color images, embed the hiding information in chromatic channels and keep the luminance unchanged.

• Support Affine transformation to deal with image distortion

• Make it a stand alone application by integrate the Matlab code with C code

Schedule/TimelineSchedule/Timeline

• Literature search and study - 2wks• Basic functionality: carrier function,

simple iconic image, basic embedding, extraction, invisible ----------- 2wks

• Embed into multiple-tile images (rotation, scaling resistant) --------------- 2wks

• Handle cropping ---------------- 1wk• Binary message and message template

function---------------------------- 1wk• Performance evaluation ------ 1wk• Wrap up and Final reports --- 1wk

Done

Done (Partial)

DoneDone

DoneDoneIn progress

Thank youThank you

• Question?