Half-Tone Watermarking Multimedia Security
Outline Half-tone technique Watermarking Method Measurement Robustness Conclusion 2
What is Half-tone? Term used in the publishing industry for a black-andwhite photograph, indicating the many shades of grey that must be reproduced In printing, a continuous tone image, such as a photograph, that has been converted into a blackand-white image Halftone images appear routinely in books, magazines, newspapers, printer outputs, and fax documents 3
Application to Halftone Fax machines ink jet printers Laser printers Magazines Newspapers All of them represent continuous tone images with small dots of one (or a few) colors of ink. 4
Example-Lena 5
Halftone technique Bit-plane 1 bit quantizer by medium value Pseudo-random-noisy binarization Dithering average dithering random dithering ordered dithering (1973) Error Diffusion (1976) 6
Average Dithering Choosing a certain constant gray level, in particular the average value of image pixels, and using it as a global threshold in deciding whether a pixel should be quantized to 0 or to 1. Simple Quantization contouring is quite perceptible. 7
Ordered Dithering It compares the pixel intensities with some pseudo-random threshold patterns or screens in order to determine its two-tone output. 0 32 8 40 2 34 10 42 48 16 56 24 50 18 58 26 12 44 4 36 14 46 6 38 60 28 42 20 62 30 44 22 3 35 11 43 1 33 9 41 51 19 59 27 49 17 57 25 15 47 7 39 13 45 5 37 63 31 45 23 61 29 43 21 0 8 2 10 0 2 3 1 12 4 14 6 3 11 1 9 15 7 13 5 8
Ordered Dithering-Example Pattern: 0 2 3 1 A pixel is 157 256/4=64 64*2 < 157 < 64*3 Halftone: 0 2 3 1 9
Random Dithering Simple and easy to implement Generate a random number 1..256; if it is greater than the image value, plot the point black a lot of "white noise", But free from "artifacts" by digital signal processing 10
Error Diffusion(1/2) It compares the sum of image pixel intensity and error from the past with a fixed threshold to determine the output The halftone error is fed forward to its adjacent neighbor using a kernel Best quality and slowest 11
Error Diffusion(2/2) 1. find the closest color available 2. Calculate the difference between the value in the image and the color you have 3. divide up these error values and distribute them over the neighboring pixels which you have not visited yet 12
Error Diffusion(2/2) 1. find the closest color available A pixel is 157. Xij = vij =157 157 > 128(Threshold) bi,j = 128 13
Error Diffusion(2/2) 2. Calculate the difference between the value in the image and the color you have 157-128 = 29 14
Error Diffusion(2/2) 3. divide up these error values and distribute them over the neighboring pixels which you have not visited yet 29*(7/48) add to next pixel 15
Performance Gray Lever Average Random Order Diffusion 16
Error Diffusion-Example(1/2) 112 0 0 112 0 0 0 157 34 0 128 34 232 121 21 232 121 21 1. 128 < 157 <256 the closest color is 128 2. Calculate the difference between the value in the image and the color you have 157-128 128 = 29 17
Error Diffusion-Example(2/2) 3. divide up these error values and distribute them over the neighboring pixels which you have not visited yet 112 0 0 0 128 47 237 130 23 112 0 0 29* 7/16 + 0 128 34 3/16 5/16 1/16 232 121 21 18
Halftone vs. watermarking Copyright, data, or time Although there are many existing image data hiding and watermarking methods, most are designed for multi-tone images and could not be applied to halftone image directly 19
Classification Pixel-wise: Change the values of (usually pseudo-randomly chosen) individual pixels (Fu and Au, 2000; Fu and Au, 2002a;) Block-wise: Divide the host image into blocks and modify some characteristic of each block. Alternating between two different dithering matrices to halftone a image (Baharav and Shaked, 1998; Hel-Or, 2001; Pei and Guo, 2003) 20
Data Hiding Watermarking for Halftone Images Author: Ming Sun Fu, Oscar C. Au Source: IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL.11, NO.4, APRIL 2002
Realignment Error correction code To help realignment: the 4 corner pixels are forced to be zero A predefined synchronization codeword embed in the center portion of the image 22
M.F. Fu & Oscar C. Au Data hiding without original multitone image Data Hiding Self Toggling (DHST) Data Hiding Pair Toggling (DHPT) Data Hiding Smart Pair Toggling (DHSPT) Data hiding with original multitone image Data Hiding Error Diffusion (DHED) Modified Data Hiding Error Diffusion (MDHED) 23
Data Hiding - Self Toggling(1/2) Encoder 1. pseudo-random number generator with a known seed decide where to be embedded 2. Force the pixel at the location to be either 0 or 255 Decoder The same random number generator with the same seed is used to identify the location 24
Data Hiding - Self Toggling(2/2) Extremely simple Low perceptual quality Salt-and pepper noise A clusters of connected pixels may be formed 0 2 3 1 25
Data Hiding Pair Toggling Master pixel: A pixel at a pseudo-random location needs to selftoggle Slave pixel: One of the neighborhood pixels is chosen randomly to self-toggle Still large salt-and-pepper noise DHST DHPT 26
Data Hiding by Smart Pair Toggling(1/4) Minimizes the connectedness of the cluster DHSPT is the same as DHPT except that the choice of the slave pixel for complementary toggling A function to Measure how connected the pixel at (m, n) is with neighboring pixels of the same color 27
Data Hiding by Smart Pair Toggling(2/4) con( m, n) f ( x, y) = = 1 0 8 i= 1 w( i) x, ) 0 Where w(i) = 1 for i = 1, 3, 6, 8 w(i) = 2 for i = 2, 4, 5, 7,, x x f ( = y y x i x 1 x 2 x 3 x 4 x 0 x 5 x 6 x 7 x 8 1 2 1 2 x 0 2 1 2 1 The con(m,n) ) is a measure of how connected the pixel at (m, n) is with neighboring pixels of the same color 28
Data Hiding by Smart Pair Toggling(3/4) Maximum con(m,n) = 12 all the same If x 0 is toggled to and the 8 neighboring are not changed, then f con = ( 0 x0, xi ) + f ( x, xi ) = 1 8 i= 1 before w( i) ( m, n) + con after x 0 ( m, n) [ ] 8 f ( x, x ) + ( = i f x0, xi ) 0 w( i) = 12 con before (m,n) con after (m,n) i= 1 -the connection before and after toggling 29
Data Hiding by Smart Pair Toggling(4/4) The slave candidate with minimum conafter(m,n) is chosen f ( x 0, xi ) + f ( x0, xi ) = 0 xi x0 Horizontal or Vertical con bef (m,n)+con aft (m,n)=10 Otherwise -con bef (m,n)+con aft (m,n)=11 con DHSPT after ( m, n) 10 con = 11 con DHSPT before DHSPT before ( m, n) ( m, n) veritcal otherwise horizontal 30
M.F. Fu & Oscar C. Au Data hiding without original multi-tone image Data Hiding Self Toggling Data Hiding Pair Toggling Data Hiding Smart Pair Toggling Data hiding with original multi-tone image Data Hiding Error Diffusion Modified Data Hiding Error Diffusion 31
Data hiding with original multi-tone image Requirement Original image is available Error diffusion Start off with DHST Using error diffusion to diffuse the selftoggling distortion to many neighboring pixels to achieve higher visual quality 32
DHED and MDHED DHED DHST is first applied followed by regular error diffusion After DHST, error diffusion is applied to the rest of the pixels MDHED The error diffusion is modified to become non-causal such that the error is fed not only to future pixels but also to past pixels 33
34 Data Hiding Error Diffusion [ ] ), ( ), ( ), ( 128 ), (, 255 128 ), (, 0 ), ( ), '( ), ( ), ( 1), ( 7 1) 1, ( 3 ) 1, ( 5 1) 1, ( 16 1 ), '( j i b j i v j i e j i v j i v j i b j i x j i x j j v j i e j i e j i e j i e j i x = < = + = + + + + = 7/16 3/16 5/16 1/16
quality measure Subjective: Typically, halftone quality is measures by just eyeballing the original and halftone Objective The human visual system acts like a lowpass filter MPSNR 35
Modified PSNR(1/2) Attempt to model the human visual system 1. A simple inverse halftone H low is generated with a low pass filter (EX:5*5 Gaussian low-pass filter) Original Image Normal PSNR computation Halftone Image 1 273 1 4 7 4 1 4 16 26 16 4 7 26 41 26 7 4 16 26 16 4 1 4 7 4 1 36
Modified PSNR(2/2) 2. H low is then fed into the regular PSNR function to generate the MPSNR O(x,y):the original grayscale image H low :halftone image though low-pass filter SxSy:image size 2 255 MPSNR =10log MSE [ H ( x, y) O( x, y) ] low MSE = S x S y 2 37
An Extension of DHSPT to color Half-tone Images Author: Mitsuji, Masayuki Inoue and Yoko Kitamura Source: International Symposium on Communications and Information Technologies 2004(ISCIT 2004)
An Extension of DHSPT to color - 2004 Extend to color images Luminance components are considered into the objective function for deciding the toggling position for visual compensation A simple extension RGB(255,0,255) (0,255,0) Low image quality 39
Color DHSPT- Proposed Method(1/3) Halftoning DHST DHSPT Original Image Halftone Image Embeddd Image YCbCr transform Compensated Image Y component y component Y-y 40
Color DHSPT- Proposed Method(2/3) 1. The embedding positions are decided by the random number and their color components are toggled 2. Only Y component is considered, because of its significance for human visual nature 3. If the Y difference between original and toggled image is small, the effect of toggling is inconspicuous D m, n = Ym, n ym, n 41
Color DHSPT- Proposed Method(3/3) 4. The evaluation measure by DHSPT for each component in 3*3 neighbor 5. These measures are combined for the evaluation of appropriate position for toggling as a pair and they given as R G B m, n m, n m, n = = = D D D m, n m, n m, n + + + D D D m, n m, n m, n con con con R G B ( m, n) ( m, n) ( m, n) 6. Find the minimum con DHSPT after 10 con = 11 con ( m, n) DHSPT before DHSPT before D m, n = Ym, n ym, n ( m, n) ( m, n) veritcal horizontal otherwise 42
Improved Techniques for Watermarking Halftone Images Author: Phil Sherry and Andreas Savakis Source:ICASSP-2004
Proposed improvement Data Hiding Cell Parity (DHCP) Data Hiding Mask Toggling (DHMT) Dispersed Pseudorandom Generator (DPRG) 44
Data Hiding Cell Parity Encoding the data stream in the parity domain instead of individual pixel Parity is contained within the 2*2 area If embedded bit is equal to parity, no action is performed. Otherwise, a complementary toggle is performed T T T T T P P T T P P T T T T T Parity ( ) mod2 = P The main motivation to increase DHSPT candidate search space (8 20) 45
Data Hiding Mask Toggling A post-halftoned watermark that works by encoding two bits in the 2*2 area Each 2-bit data is assigned a set of five possible encoding masks The masks maintain the cell intensity and minimizing visual degradation 46
High-Capacity Data Hiding in Halftone Images Using Minimal-Error Bit Searching and Least-Mean Square Filter Author: Soo-Chang Pei, Jing-Ming Guo Source: IEEE Transactions on Image Processing, VOL. 15, NO.6,June 2006
Stand EDF flow chat 1. find the closest color available 2. Calculate the difference between the value in the image and the color you have 3. divide up these error values and distribute them over the neighboring pixels which you have not visited yet 48
3 bit gray code and information bit 49
Generate the Gray code Generate the 3-bit Gray code and address each codeword with a corresponding information bit 0 or 1 The information bits 0 and 1 are arranged alternately, and the table is divided into two groups 50
Hiding Data If the information bit 0 is embedded into the host EDF images, then we check if the binary output vector B i,j is mapped to 0-group If yes, then no bit in this vector should be varied if no, then we just need to modify one bit in this vector 51
Proposed high-capacity MEBS data-hiding encoder 52
Minimal Error Bit Searching(1/2) Encoder Be used for judging which bit in the vector is the most suitable candidate 1. Find the complementary binary vector B i, j of the temporary binary output vector B i, j { } k 1,, = k k if bi j bi, j k = 1,2,..., n where bi, j = b = 1 = k if i, j 0, B i, j 2. Calculate all the differences between the k gray outputs v i, j and the complementary binary outputs, where k=1,2,..,n k k k e i, j = vi, j bi, j 0 53
Minimal Error Bit Searching(2/2) { } m k 3. Suppose ei, j = min ei, j k = 1,2,..., n ;then, the modified binary output vector is formed as { } 1 2 m n b, b,..., b b ' B i, j = i, j i, j i, j,..., i, j Decoder: Using the table lookup method (LUT) Ex:from the gray code table, {1,0,1} represents 1 had been embedded 54
Modified MEBS When the capacity gets as high as 50%, the watermarked image quality will be degraded k k k e i, j = vi, j bi, j e k k ERRth, if ERRth < vi, j bi, j The best choice is 135 k = ERRth if ERRth > vi, j b k k vi, j bi, j otherwise k i, j, k i, j 55
Robust Watermarking Divide each original gray-level images of size into several cells of size In each cell, embed the same information bit Use the majority voting to recover the damaged portion caused by all kinds of distortions 56
Pei s Measure For an image with size P*Q W m,n :the human visual system (HVS) coefficient at position (m,n) X i,j :the original image B i,j :the halftone image PSNR = P Q x 2 P* Q*255 i, j i= 1 j= 1 m, n R w m, n b i+ m, j+ n LMS s distribution 57
Result In this paper, a high-capacity data-hiding method in halftone image using the MEBS technique is proposed From the experimental results, they prove that, when the capacity is as high as 33.33%, good qualities of the embedded halftone images can be achieved 58
Conclusion Halftone is still dominant in usual office challenge Physical challenges:scanner resolution Affine Transformation:Rotation Translation Dirt and Stains.. 59
Reference S.F. Ming and C.A. Oscar, Data Hiding Watermarking for Halftone Image, IEEE Trans. on Image Processing, April 2002 Mitsuji Muneyasu, Masayuki Inoue and Yoko Kitamura, An Extension of DHSPT to Color Half-tone Images, ISCIT 2004 Phil Sherry and Andreas Savakis, Improved Techniques for Watermarking Halftone Images, ICASSP 2004 Kimiyoshi Arishima, Makoto Kitamura & Fumitaka Ono, A study on the data hiding rate for halftone images, Signal Processing and Its Applications, August 2005 S. C. Pei and J. M. Guo, "High-capacity data hiding in halftone images using minimal error bit searching," IEEE International Conference on Image Processing, Singapore, Oct 2004 J. M. Guo, S. C. Pei, and H. Lee, "Watermarking in halftone images with paritymatched error diffusion," IEEE International Conference on Speech, Acoustics, and Signal Processing, U.S.A. March 2005 S. C. Pei, and J. M. Guo, High capacity data hiding in halftone images using minimal error bit searching and least mean square filter, IEEE Trans. Image Processing, vol. 15, no. 6, pp. 1665-1679, June, 2006 60