The Choice of Filter Banks for Wavelet-based Robust Digital Watermarking p. 1/18

The Choice of Filter Banks for Wavelet-based Robust Digital Watermarking Martin Dietze martin.dietze@buckingham.ac.uk Sabah Jassim sabah.jassim@buckingham.ac.uk The University of Buckingham United Kingdom The Choice of Filter Banks for Wavelet-based Robust Digital Watermarking p. 1/18

Buckingham Overview Motivation Watermarking in the DWT-Domain Watermarking Performance Factors Experimental Setup The Quality Measurement Results Conclusions and Further Work The Choice of Filter Banks for Wavelet-based Robust Digital Watermarking p. 2/18

Motivation Watermarking in the Wavelet Domain: Different Filters The Choice of Filter Banks for Wavelet-based Robust Digital Watermarking p. 3/18

Motivation Watermarking in the Wavelet Domain: Different Filters The choice of filter for Watermarking: Known to be important from compression applications Expected to influence watermarking performance Possible optimization for robustness or image quality The Choice of Filter Banks for Wavelet-based Robust Digital Watermarking p. 3/18

Motivation Watermarking in the Wavelet Domain: Different Filters The choice of filter for Watermarking: Known to be important from compression applications Expected to influence watermarking performance Possible optimization for robustness or image quality Yet little is said in the literature on the filter choice The Choice of Filter Banks for Wavelet-based Robust Digital Watermarking p. 3/18

Buckingham Motivation Watermarking in the Wavelet Domain: Different Filters The choice of filter for Watermarking: Known to be important from compression applications Expected to influence watermarking performance Possible optimization for robustness or image quality Yet little is said in the literature on the filter choice This paper aims to rank filters by WM requirements The Choice of Filter Banks for Wavelet-based Robust Digital Watermarking p. 3/18

Watermarking in the DWT Domain Marking Process: Decompose image The Choice of Filter Banks for Wavelet-based Robust Digital Watermarking p. 4/18

Watermarking in the DWT Domain Marking Process: Decompose image... up to depth The Choice of Filter Banks for Wavelet-based Robust Digital Watermarking p. 4/18

Watermarking in the DWT Domain Marking Process: Decompose image... up to depth Choose subbands The Choice of Filter Banks for Wavelet-based Robust Digital Watermarking p. 4/18

Watermarking in the DWT Domain Marking Process: Decompose image... up to depth Choose subbands Modify coefficients The Choice of Filter Banks for Wavelet-based Robust Digital Watermarking p. 4/18

Watermarking in the DWT Domain Marking Process: Decompose image... up to depth Choose subbands Modify coefficients... in secret places The Choice of Filter Banks for Wavelet-based Robust Digital Watermarking p. 4/18

Watermarking in the DWT Domain Marking Process: Decompose image... up to depth Choose subbands Modify coefficients... in secret places Finally apply the inverse transform for marked image The Choice of Filter Banks for Wavelet-based Robust Digital Watermarking p. 4/18

Watermarking Performance Factors The Choice of Filter Banks for Wavelet-based Robust Digital Watermarking p. 5/18

Watermarking Performance Factors Factors influencing a DWT-based scheme s performance: Choice of Filter Subband depth of marking Decomposition Scheme for embedding The Choice of Filter Banks for Wavelet-based Robust Digital Watermarking p. 5/18

Watermarking Performance Factors Factors influencing a DWT-based scheme s performance: Choice of Filter Subband depth of marking Decomposition Scheme for embedding Factors shared with non-dwt schemes: Embedding technique Embedding intensity Image properties (e.g. variation of texture) The Choice of Filter Banks for Wavelet-based Robust Digital Watermarking p. 5/18

Experimental Setup The Filters tested are: Orthogonal Filters: Biorthogonal Filters: Name Length Name Length Haar 2 Daub4 4 Daub6 6 Daub8 8 Antonini 7/9 Brislawn 9/7 Odegard 9/7 Villa1 9/7 Villa2 13/11 Villa3 6/10 Villa4 5/3 Villa5 2/6 Villa6 9/3 The Choice of Filter Banks for Wavelet-based Robust Digital Watermarking p. 6/18

Experimental Setup (cont.) The Watermarking Technique for Testing: The Choice of Filter Banks for Wavelet-based Robust Digital Watermarking p. 7/18

Buckingham Experimental Setup (cont.) The Watermarking Technique for Testing: Mark the chosen subbands maximum DWT coefficients using non-blind multiplicative embedding: A Pyramid decomposition is used Subbands for marking: 1, 1-2, 1-3 or 1-4 The Choice of Filter Banks for Wavelet-based Robust Digital Watermarking p. 7/18

Buckingham Experimental Setup (cont.) The Watermarking Technique for Testing: Mark the chosen subbands maximum DWT coefficients using non-blind multiplicative embedding: A Pyramid decomposition is used Subbands for marking: 1, 1-2, 1-3 or 1-4 Marking intensity: 20%-80% The Choice of Filter Banks for Wavelet-based Robust Digital Watermarking p. 7/18

Buckingham Experimental Setup (cont.) The Watermarking Technique for Testing: Mark the chosen subbands maximum DWT coefficients using non-blind multiplicative embedding: A Pyramid decomposition is used Subbands for marking: 1, 1-2, 1-3 or 1-4 Marking intensity: 20%-80% The watermark can be any file; here: binary image The Choice of Filter Banks for Wavelet-based Robust Digital Watermarking p. 7/18

Buckingham Experimental Setup (cont.) The Watermarking Technique for Testing: Mark the chosen subbands maximum DWT coefficients using non-blind multiplicative embedding: A Pyramid decomposition is used Subbands for marking: 1, 1-2, 1-3 or 1-4 Marking intensity: 20%-80% The watermark can be any file; here: binary image Use the embedding coordinates for reading the mark The Choice of Filter Banks for Wavelet-based Robust Digital Watermarking p. 7/18

Experimental Setup (cont.) Watermark Embedding Watermark (binary image) Cover Image The Choice of Filter Banks for Wavelet-based Robust Digital Watermarking p. 8/18

Experimental Setup (cont.) Watermark Embedding Image Quality Scoring (MSE) Watermark (binary image) Cover Image The Choice of Filter Banks for Wavelet-based Robust Digital Watermarking p. 8/18

Experimental Setup (cont.) Watermark Embedding Image Quality Scoring (MSE) Compression Attack (JPG or DWT) Watermark (binary image) Cover Image The Choice of Filter Banks for Wavelet-based Robust Digital Watermarking p. 8/18

Experimental Setup (cont.) Watermark Embedding Image Quality Scoring (MSE) Compression Attack (JPG or DWT) Watermark Detection Watermark (binary image) Detected Watermark Cover Image The Choice of Filter Banks for Wavelet-based Robust Digital Watermarking p. 8/18

Experimental Setup (cont.) Watermark Embedding Image Quality Scoring (MSE) Compression Attack (JPG or DWT) Watermark Detection Detection Scoring (MSE and L^qd) Watermark (binary image) Detected Watermark Cover Image The Choice of Filter Banks for Wavelet-based Robust Digital Watermarking p. 8/18

The Quality Measurement Our watermark is an image how can we measure detection quality automatically? The Choice of Filter Banks for Wavelet-based Robust Digital Watermarking p. 9/18

The Quality Measurement Our watermark is an image how can we measure detection quality automatically? Humans can still identify a heavily distorted logo The Choice of Filter Banks for Wavelet-based Robust Digital Watermarking p. 9/18

The Quality Measurement Our watermark is an image how can we measure detection quality automatically? Humans can still identify a heavily distorted logo Can we mimic this in software? The Choice of Filter Banks for Wavelet-based Robust Digital Watermarking p. 9/18

Buckingham The Quality Measurement Our watermark is an image how can we measure detection quality automatically? Humans can still identify a heavily distorted logo Can we mimic this in software? The quality measurement: Is a modification of the pseudo-norm for image querying [Jacobs et.al. 1995] The Choice of Filter Banks for Wavelet-based Robust Digital Watermarking p. 9/18

Buckingham The Quality Measurement Our watermark is an image how can we measure detection quality automatically? Humans can still identify a heavily distorted logo Can we mimic this in software? The quality measurement: Is a modification of the pseudo-norm for image querying [Jacobs et.al. 1995] Exploits the DWT-domain s multiresolutional properties The Choice of Filter Banks for Wavelet-based Robust Digital Watermarking p. 9/18

Buckingham The Quality Measurement (cont.) Philosophy: Mark Recognition depends on only few coefficients The Choice of Filter Banks for Wavelet-based Robust Digital Watermarking p. 10/18

Buckingham The Quality Measurement (cont.) Philosophy: Mark Recognition depends on only few coefficients Use rough details: Coarse subbands The Choice of Filter Banks for Wavelet-based Robust Digital Watermarking p. 10/18

Buckingham The Quality Measurement (cont.) Philosophy: Mark Recognition depends on only few coefficients Use rough details: Coarse subbands Insignificant coefficients can be discarded The Choice of Filter Banks for Wavelet-based Robust Digital Watermarking p. 10/18

Buckingham The Quality Measurement (cont.) The Algorithm: Fully decompose both images (Haar, Pyramid) The Choice of Filter Banks for Wavelet-based Robust Digital Watermarking p. 11/18

Buckingham The Quality Measurement (cont.) The Algorithm: Fully decompose both images (Haar, Pyramid) Set top coefficients to 1 or -1, and all others to 0 The Choice of Filter Banks for Wavelet-based Robust Digital Watermarking p. 11/18

Buckingham The Quality Measurement (cont.) The Algorithm: Fully decompose both images (Haar, Pyramid) Set top coefficients to 1 or -1, and all others to 0 For each image s nonzero coefficients: If the other image s corresponding coefficient differs, add a value from a weight table The Choice of Filter Banks for Wavelet-based Robust Digital Watermarking p. 11/18

Buckingham The Quality Measurement (cont.) The Algorithm: Fully decompose both images (Haar, Pyramid) Set top coefficients to 1 or -1, and all others to 0 For each image s nonzero coefficients: If the other image s corresponding coefficient differs, add a value from a weight table Weights are experimentally determined [Jacobs] The Choice of Filter Banks for Wavelet-based Robust Digital Watermarking p. 11/18

Buckingham The Quality Measurement (cont.) The Algorithm: Fully decompose both images (Haar, Pyramid) Set top coefficients to 1 or -1, and all others to 0 For each image s nonzero coefficients: If the other image s corresponding coefficient differs, add a value from a weight table Weights are experimentally determined [Jacobs] The is the two sums normalized minimum The Choice of Filter Banks for Wavelet-based Robust Digital Watermarking p. 11/18

Buckingham The Quality Measurement (cont.) The Algorithm: Fully decompose both images (Haar, Pyramid) Set top coefficients to 1 or -1, and all others to 0 For each image s nonzero coefficients: If the other image s corresponding coefficient differs, add a value from a weight table Weights are experimentally determined [Jacobs] The is the two sums normalized minimum The lower the value the better the match The Choice of Filter Banks for Wavelet-based Robust Digital Watermarking p. 11/18

Results Testing parameters: The Choice of Filter Banks for Wavelet-based Robust Digital Watermarking p. 12/18

Results Testing parameters: Watermark intensity The Choice of Filter Banks for Wavelet-based Robust Digital Watermarking p. 12/18

Results Testing parameters: Watermark intensity Attack compression ratio The Choice of Filter Banks for Wavelet-based Robust Digital Watermarking p. 12/18

Results Testing parameters: Watermark intensity Attack compression ratio Kind of attack (JPG or DWT) The Choice of Filter Banks for Wavelet-based Robust Digital Watermarking p. 12/18

Results Testing parameters: Watermark intensity Attack compression ratio Kind of attack (JPG or DWT) The image characteristics The Choice of Filter Banks for Wavelet-based Robust Digital Watermarking p. 12/18

Results Testing parameters: Watermark intensity Attack compression ratio Kind of attack (JPG or DWT) The image characteristics Chosen subband depth for marking The Choice of Filter Banks for Wavelet-based Robust Digital Watermarking p. 12/18

Results Testing parameters: Watermark intensity Attack compression ratio Kind of attack (JPG or DWT) The image characteristics Chosen subband depth for marking Rankings depend most on image, subband depth and kind of attack. The Choice of Filter Banks for Wavelet-based Robust Digital Watermarking p. 12/18

Results (cont.) Image Degradation: The Choice of Filter Banks for Wavelet-based Robust Digital Watermarking p. 13/18

Results (cont.) Image Degradation: Marking only subband 1: little degradation The Choice of Filter Banks for Wavelet-based Robust Digital Watermarking p. 13/18

Results (cont.) Image Degradation: Marking only subband 1: little degradation Increasing the subband depth: visible artifacts The Choice of Filter Banks for Wavelet-based Robust Digital Watermarking p. 13/18

Results (cont.) Image Degradation: Marking only subband 1: little degradation Increasing the subband depth: visible artifacts Reasons: Choice of significant coefficients for marking and 1:4 relationship from coarser to finer subbands. The Choice of Filter Banks for Wavelet-based Robust Digital Watermarking p. 13/18

Buckingham Results (cont.) Image Degradation: Marking only subband 1: little degradation Increasing the subband depth: visible artifacts Reasons: Choice of significant coefficients for marking and 1:4 relationship from coarser to finer subbands. However: Much better robustness at subband depths The Choice of Filter Banks for Wavelet-based Robust Digital Watermarking p. 13/18

Results (cont.) The Degradation Rankings: Degradation against subbands measured in MSE Antonini, Brislawn, Villa1 Haar, Daub4, Daub6, Daub8 Odegard, Villa2 Villa3 Villa4 Villa5 Villa6 Rank 1 2 3 4 5 6 7 1 2 3 4 average Subbands The Choice of Filter Banks for Wavelet-based Robust Digital Watermarking p. 14/18

Results (cont.) The Detection Rankings (JPG attack): Detection (JPG attack) against subbands measured in L^qd Rank 1 2 3 4 5 6 7 8 Haar, Daub4, Daub6, Daub8 Antonini, Brislawn, Villa1 Villa2 Villa3 Villa4 Villa5 Villa6 Odegard 1 2 3 4 average Subbands The Choice of Filter Banks for Wavelet-based Robust Digital Watermarking p. 15/18

Results (cont.) The Detection Rankings (DWT attack): Detection (DWT attack) against subbands measured in L^qd Rank 1 2 3 4 5 6 7 8 9 10 11 Haar Antonini, Brislawn, Villa1 Villa2 Villa3 Villa4 Villa5 Villa6 Odegard Daub4 Daub6 Daub8 1 2 3 4 average Subbands The Choice of Filter Banks for Wavelet-based Robust Digital Watermarking p. 16/18

Conclusions Quality and robustness against attacks are contradicting requirements The Choice of Filter Banks for Wavelet-based Robust Digital Watermarking p. 17/18

Conclusions Quality and robustness against attacks are contradicting requirements In both cases, performance of filters depends on the subband depth (2 is a reasonable compromise) Adapting the marking intensity leads to a better tradeoff The Choice of Filter Banks for Wavelet-based Robust Digital Watermarking p. 17/18

Conclusions Quality and robustness against attacks are contradicting requirements In both cases, performance of filters depends on the subband depth (2 is a reasonable compromise) Adapting the marking intensity leads to a better tradeoff Other influences are beyond the marker s control The Choice of Filter Banks for Wavelet-based Robust Digital Watermarking p. 17/18

Conclusions Quality and robustness against attacks are contradicting requirements In both cases, performance of filters depends on the subband depth (2 is a reasonable compromise) Adapting the marking intensity leads to a better tradeoff Other influences are beyond the marker s control Villa3 has good overall properties (followed by the group around Antonini and Villa6) The Choice of Filter Banks for Wavelet-based Robust Digital Watermarking p. 17/18

Further Work This work has been extended to use the SCS (Scalar Costa Scheme, introduced by Eggers in 2000) instead of the Multiplicative embedding. The Choice of Filter Banks for Wavelet-based Robust Digital Watermarking p. 18/18

Further Work This work has been extended to use the SCS (Scalar Costa Scheme, introduced by Eggers in 2000) instead of the Multiplicative embedding. This work is part of an ongoing project on Wavelet-based Second Generation Watermarking. The Choice of Filter Banks for Wavelet-based Robust Digital Watermarking p. 18/18

Further Work This work has been extended to use the SCS (Scalar Costa Scheme, introduced by Eggers in 2000) instead of the Multiplicative embedding. This work is part of an ongoing project on Wavelet-based Second Generation Watermarking. For more details, including a more detailed version of the paper, please see my web pages on the project: http://herbert.the-little-red-haired-girl.org/en/research/ The Choice of Filter Banks for Wavelet-based Robust Digital Watermarking p. 18/18