![Page 1: An Improved Method Of Content Based Image Watermarking Arvind Kumar Parthasarathy and Subhash Kak 黃阡廷 2008/12/3](https://reader036.vdocuments.mx/reader036/viewer/2022062315/5697bf8a1a28abf838c8a8ae/html5/thumbnails/1.jpg)
An Improved Method Of Content BasedImage WatermarkingArvind Kumar Parthasarathy and Subhash Kak
黃阡廷2008/12/3
![Page 2: An Improved Method Of Content Based Image Watermarking Arvind Kumar Parthasarathy and Subhash Kak 黃阡廷 2008/12/3](https://reader036.vdocuments.mx/reader036/viewer/2022062315/5697bf8a1a28abf838c8a8ae/html5/thumbnails/2.jpg)
Author
Arvind Kumar ParthasarathyReceived the M.Sc. degree in electrical engineering from Louisiana State University, Baton Rouge, Louisiana in 2006His researchinterests include digital image watermarking, imageprocessing, cryptography and network security
SubhashKakTheDonald C. and Elaine T.DelauneDistinguished Professor of Electrical Engineering atLouisiana State University at Baton Rouge
Manuscript received September 8, 2006; revised February 21, 2007.
IEEE TRANSACTIONS ON BROADCASTING, VOL. 53, NO. 2, JUNE 2007
![Page 3: An Improved Method Of Content Based Image Watermarking Arvind Kumar Parthasarathy and Subhash Kak 黃阡廷 2008/12/3](https://reader036.vdocuments.mx/reader036/viewer/2022062315/5697bf8a1a28abf838c8a8ae/html5/thumbnails/3.jpg)
OUTLINE
FREQUENCY-BASED WATERMARKING1
CORRELATION-BASED WATERMARKING2
PROPOSED SCHEME3
WATERMARK EVALUATION4
EXPERIMENTAL RESULTS5
ATTACKS AND ANALYSIS OF RESULTS6
CONCLUSIONS7
![Page 4: An Improved Method Of Content Based Image Watermarking Arvind Kumar Parthasarathy and Subhash Kak 黃阡廷 2008/12/3](https://reader036.vdocuments.mx/reader036/viewer/2022062315/5697bf8a1a28abf838c8a8ae/html5/thumbnails/4.jpg)
FREQUENCY-BASED WATERMARKING
Addition of the watermark is done in a transformed domain, DCT and DWT are two such popular transforms.
Frequency-based techniques are very robust against attacks involving image compression and filtering.
![Page 5: An Improved Method Of Content Based Image Watermarking Arvind Kumar Parthasarathy and Subhash Kak 黃阡廷 2008/12/3](https://reader036.vdocuments.mx/reader036/viewer/2022062315/5697bf8a1a28abf838c8a8ae/html5/thumbnails/5.jpg)
FREQUENCY-BASED WATERMARKING
Watermarking in the DCT domain is usually performed on the lower or the mid-band frequencies, as higher frequencies are lost when the image is compressed.
Original Image DCT TransformedEmbeddedWatermark
DCT Inverse-Transformed
![Page 6: An Improved Method Of Content Based Image Watermarking Arvind Kumar Parthasarathy and Subhash Kak 黃阡廷 2008/12/3](https://reader036.vdocuments.mx/reader036/viewer/2022062315/5697bf8a1a28abf838c8a8ae/html5/thumbnails/6.jpg)
FREQUENCY-BASED WATERMARKING
This paper proposes a robust and transparent scheme of watermarking.
We implement changes in this algorithm without much distortion.
![Page 7: An Improved Method Of Content Based Image Watermarking Arvind Kumar Parthasarathy and Subhash Kak 黃阡廷 2008/12/3](https://reader036.vdocuments.mx/reader036/viewer/2022062315/5697bf8a1a28abf838c8a8ae/html5/thumbnails/7.jpg)
CORRELATION-BASED WATERMARKINK
In most schemes, the watermark is typically a pseudo randomly generated noise sequence.
The generalized algorithm of most correlation-based spread spectrum watermarking in a spatial domain is based on the following equation:
),(),(),( jiWkjiIjiWI
![Page 8: An Improved Method Of Content Based Image Watermarking Arvind Kumar Parthasarathy and Subhash Kak 黃阡廷 2008/12/3](https://reader036.vdocuments.mx/reader036/viewer/2022062315/5697bf8a1a28abf838c8a8ae/html5/thumbnails/8.jpg)
CORRELATION-BASED WATERMARKINK
We will consider an invisible watermarking method that is capable of hiding the watermark information in the cover image in an unnoticeable.
Our watermarking scheme deals with the extraction of the watermark information in the absence of the original image.
![Page 9: An Improved Method Of Content Based Image Watermarking Arvind Kumar Parthasarathy and Subhash Kak 黃阡廷 2008/12/3](https://reader036.vdocuments.mx/reader036/viewer/2022062315/5697bf8a1a28abf838c8a8ae/html5/thumbnails/9.jpg)
PROPOSED SCHEME
We divide our scheme into three steps1. Generation of a mask based on the
perceptual properties of the image
2. Watermarking, by spreading the d-sequence in the frequency domain, by multiplying it with the weights calculated from step 1
3. Extraction of the watermark by using a correlation-based method.
![Page 10: An Improved Method Of Content Based Image Watermarking Arvind Kumar Parthasarathy and Subhash Kak 黃阡廷 2008/12/3](https://reader036.vdocuments.mx/reader036/viewer/2022062315/5697bf8a1a28abf838c8a8ae/html5/thumbnails/10.jpg)
PROPOSED SCHEME
A. Just Noticeable Distortion (JND) Visual Mask
JND is defined a measure referring to the capability of a human observer to detect noise or distortion in the field of view.
A good JND mask would depend on the accurate extraction of the luminance.
![Page 11: An Improved Method Of Content Based Image Watermarking Arvind Kumar Parthasarathy and Subhash Kak 黃阡廷 2008/12/3](https://reader036.vdocuments.mx/reader036/viewer/2022062315/5697bf8a1a28abf838c8a8ae/html5/thumbnails/11.jpg)
PROPOSED SCHEME
Our scheme is image adaptive as it incorporates the local information extracted from the image.
The algorithm that is used to extract the DCT coefficients Image is segmented into non-overlapping
blocks of size 8*8.
1
0
1
0
8,0 where),,(),(
N
n
N
nnn jijifByxf
![Page 12: An Improved Method Of Content Based Image Watermarking Arvind Kumar Parthasarathy and Subhash Kak 黃阡廷 2008/12/3](https://reader036.vdocuments.mx/reader036/viewer/2022062315/5697bf8a1a28abf838c8a8ae/html5/thumbnails/12.jpg)
PROPOSED SCHEME
Perform the DCT by MATLAB
The DC coefficient is proportional to the average pixel value
AC coefficient describe their variation around the DC
![Page 13: An Improved Method Of Content Based Image Watermarking Arvind Kumar Parthasarathy and Subhash Kak 黃阡廷 2008/12/3](https://reader036.vdocuments.mx/reader036/viewer/2022062315/5697bf8a1a28abf838c8a8ae/html5/thumbnails/13.jpg)
PROPOSED SCHEME
Texture: It is define as the quality of the object
To determine a measure for the texture information within each block based on the energy in the ac coefficients
)log( 20
63
1
2 vvPi
iT 63...1,0, iviwhere are the 64 DCT
coefficients of the 8*8 block
![Page 14: An Improved Method Of Content Based Image Watermarking Arvind Kumar Parthasarathy and Subhash Kak 黃阡廷 2008/12/3](https://reader036.vdocuments.mx/reader036/viewer/2022062315/5697bf8a1a28abf838c8a8ae/html5/thumbnails/14.jpg)
PROPOSED SCHEME
Normalized values are assigned to the corresponding block.
For an image matrix of size 512*512 we will have a matrix of size 64*64 where each one of those value corresponds to the texture information of each 8*8 block
tP
)max(
64
T
TT P
PM
![Page 15: An Improved Method Of Content Based Image Watermarking Arvind Kumar Parthasarathy and Subhash Kak 黃阡廷 2008/12/3](https://reader036.vdocuments.mx/reader036/viewer/2022062315/5697bf8a1a28abf838c8a8ae/html5/thumbnails/15.jpg)
PROPOSED SCHEME
Edge: Edges are extracted from the pixel domain and this information is useful in determining the amount of watermark information.
We use two methods to extract edge
![Page 16: An Improved Method Of Content Based Image Watermarking Arvind Kumar Parthasarathy and Subhash Kak 黃阡廷 2008/12/3](https://reader036.vdocuments.mx/reader036/viewer/2022062315/5697bf8a1a28abf838c8a8ae/html5/thumbnails/16.jpg)
PROPOSED SCHEME
Edge extra using canny operator Edge extraction on phase congruency
![Page 17: An Improved Method Of Content Based Image Watermarking Arvind Kumar Parthasarathy and Subhash Kak 黃阡廷 2008/12/3](https://reader036.vdocuments.mx/reader036/viewer/2022062315/5697bf8a1a28abf838c8a8ae/html5/thumbnails/17.jpg)
PROPOSED SCHEME
Using a binary edge map, we calculate the normalized edge information for each block using the formula
)max(
64
E
EE P
PM
Ep is the cardinality of set of pixels at edge locations in each block
![Page 18: An Improved Method Of Content Based Image Watermarking Arvind Kumar Parthasarathy and Subhash Kak 黃阡廷 2008/12/3](https://reader036.vdocuments.mx/reader036/viewer/2022062315/5697bf8a1a28abf838c8a8ae/html5/thumbnails/18.jpg)
PROPOSED SCHEME
Corner: A corner represents the point where two edges meet and the human is more sensitive to changes made in these places.
We make use of an improved corner detection algorithm based on curvature scale space (CSS).
![Page 19: An Improved Method Of Content Based Image Watermarking Arvind Kumar Parthasarathy and Subhash Kak 黃阡廷 2008/12/3](https://reader036.vdocuments.mx/reader036/viewer/2022062315/5697bf8a1a28abf838c8a8ae/html5/thumbnails/19.jpg)
PROPOSED SCHEME
The main steps:1. Extracting the edge information/contours
from a binary edge map.2. Filling in the gaps in the contours.3. Computing the curvature at a fixed low
scale to retain all the true corners4. The curvature local maxima are
considered as corners while eliminating the rounded and false corners resulting from noise using adaptive threshold and the angle of corner
![Page 20: An Improved Method Of Content Based Image Watermarking Arvind Kumar Parthasarathy and Subhash Kak 黃阡廷 2008/12/3](https://reader036.vdocuments.mx/reader036/viewer/2022062315/5697bf8a1a28abf838c8a8ae/html5/thumbnails/20.jpg)
PROPOSED SCHEME
The effective detection of corners for a Lena image based on the curvature scale space
![Page 21: An Improved Method Of Content Based Image Watermarking Arvind Kumar Parthasarathy and Subhash Kak 黃阡廷 2008/12/3](https://reader036.vdocuments.mx/reader036/viewer/2022062315/5697bf8a1a28abf838c8a8ae/html5/thumbnails/21.jpg)
PROPOSED SCHEME
On obtaining the corners by the above method we calculate the corner information for each block of the image using the formula
)max(
64
C
CC P
PM
CP is the cardinality of the group of pixels determined to be corners
![Page 22: An Improved Method Of Content Based Image Watermarking Arvind Kumar Parthasarathy and Subhash Kak 黃阡廷 2008/12/3](https://reader036.vdocuments.mx/reader036/viewer/2022062315/5697bf8a1a28abf838c8a8ae/html5/thumbnails/22.jpg)
PROPOSED SCHEME
Luminance: It is defined as the way the human eye perceives brightness of different colors.
![Page 23: An Improved Method Of Content Based Image Watermarking Arvind Kumar Parthasarathy and Subhash Kak 黃阡廷 2008/12/3](https://reader036.vdocuments.mx/reader036/viewer/2022062315/5697bf8a1a28abf838c8a8ae/html5/thumbnails/23.jpg)
PROPOSED SCHEME
Our scheme utilizes the luminance factor that is calculated by measuring the average pixel value of the gray scale image for that block
64L
L
PM
is the sum of all the pixel values in the blockLP
LM is the average of the luminance values within the considered block
![Page 24: An Improved Method Of Content Based Image Watermarking Arvind Kumar Parthasarathy and Subhash Kak 黃阡廷 2008/12/3](https://reader036.vdocuments.mx/reader036/viewer/2022062315/5697bf8a1a28abf838c8a8ae/html5/thumbnails/24.jpg)
PROPOSED SCHEME
After obtaining the four values corresponding to the texture, edge, corners and the luminance we generate the initial mask using the equation
CETI MMMJ 2
1
![Page 25: An Improved Method Of Content Based Image Watermarking Arvind Kumar Parthasarathy and Subhash Kak 黃阡廷 2008/12/3](https://reader036.vdocuments.mx/reader036/viewer/2022062315/5697bf8a1a28abf838c8a8ae/html5/thumbnails/25.jpg)
PROPOSED SCHEME
The human vision system is more sensitive to the changes in intensity in the mid-gray region
Hence a correction to the initial JND parameter value is introduced and the final JND parameter value for each block is calculated as
2)128( LIF MJJ
![Page 26: An Improved Method Of Content Based Image Watermarking Arvind Kumar Parthasarathy and Subhash Kak 黃阡廷 2008/12/3](https://reader036.vdocuments.mx/reader036/viewer/2022062315/5697bf8a1a28abf838c8a8ae/html5/thumbnails/26.jpg)
PROPOSED SCHEME
B. Watermark EmbeddingThe JND value controls the strength of
watermark for each block
The strength of the watermark component embedded, in a block with a low JND value
![Page 27: An Improved Method Of Content Based Image Watermarking Arvind Kumar Parthasarathy and Subhash Kak 黃阡廷 2008/12/3](https://reader036.vdocuments.mx/reader036/viewer/2022062315/5697bf8a1a28abf838c8a8ae/html5/thumbnails/27.jpg)
PROPOSED SCHEME
Embedding the watermark in the high frequency removal of the watermark through
compression and noise attacks
Embedding the watermark in the low frequency visible changes in the watermarked image
![Page 28: An Improved Method Of Content Based Image Watermarking Arvind Kumar Parthasarathy and Subhash Kak 黃阡廷 2008/12/3](https://reader036.vdocuments.mx/reader036/viewer/2022062315/5697bf8a1a28abf838c8a8ae/html5/thumbnails/28.jpg)
PROPOSED SCHEME
For each 8*8 transformed block the d-sequence multiplied by a scaling factor and the JND mask is added into the selected mid-frequency DCT components
![Page 29: An Improved Method Of Content Based Image Watermarking Arvind Kumar Parthasarathy and Subhash Kak 黃阡廷 2008/12/3](https://reader036.vdocuments.mx/reader036/viewer/2022062315/5697bf8a1a28abf838c8a8ae/html5/thumbnails/29.jpg)
PROPOSED SCHEME
The watermark embedding is done using the formula
mid
midFW FvubvuI
FvudbJbvuIbvuI
, ),,(
, ))((),,(),,(
),,( bvuIW),,( bvuI
)(bJ F
dmidF
is the modified DCT coefficient in location for block b),( vu
is the DCT coefficient in location for block b),( vuis the scaling factor
is the JND value generated for the block from the equation above
is the d-sequence generated
is the middle frequencies of the DCT block
![Page 30: An Improved Method Of Content Based Image Watermarking Arvind Kumar Parthasarathy and Subhash Kak 黃阡廷 2008/12/3](https://reader036.vdocuments.mx/reader036/viewer/2022062315/5697bf8a1a28abf838c8a8ae/html5/thumbnails/30.jpg)
PROPOSED SCHEME
Finally, the block containing the watermarked DCT coefficients is inverse-transformed to obtain the final watermarked image.
Performed inverse-transformed by MATLAB
![Page 31: An Improved Method Of Content Based Image Watermarking Arvind Kumar Parthasarathy and Subhash Kak 黃阡廷 2008/12/3](https://reader036.vdocuments.mx/reader036/viewer/2022062315/5697bf8a1a28abf838c8a8ae/html5/thumbnails/31.jpg)
PROPOSED SCHEME
C. Watermark DetectionThe image is first broken down into the
same 8*8 blocks
The DCT coefficients of the mid-frequency values thus obtained are compared with the d-sequence
![Page 32: An Improved Method Of Content Based Image Watermarking Arvind Kumar Parthasarathy and Subhash Kak 黃阡廷 2008/12/3](https://reader036.vdocuments.mx/reader036/viewer/2022062315/5697bf8a1a28abf838c8a8ae/html5/thumbnails/32.jpg)
PROPOSED SCHEME
TbCif
TbCif
bWbIN
bC
)( 1
)( 0bit watermarkRecovered
))()((1
)(n Correlatio *
T)(bC
)(* bI
)(bW
is the threshold level
is the correlation value for block b
is the DCT coefficient of the watermarked image
is the d-sequence that is generated using the same prime number
![Page 33: An Improved Method Of Content Based Image Watermarking Arvind Kumar Parthasarathy and Subhash Kak 黃阡廷 2008/12/3](https://reader036.vdocuments.mx/reader036/viewer/2022062315/5697bf8a1a28abf838c8a8ae/html5/thumbnails/33.jpg)
WATERMARK EVALUATION
Signal to noise ratio (SNR) effectively measures the quality of the watermarked image as compared to the original image.
The larger the value of e(x,y) the greater
),(),(),( yxIyxIyxe W
![Page 34: An Improved Method Of Content Based Image Watermarking Arvind Kumar Parthasarathy and Subhash Kak 黃阡廷 2008/12/3](https://reader036.vdocuments.mx/reader036/viewer/2022062315/5697bf8a1a28abf838c8a8ae/html5/thumbnails/34.jpg)
WATERMARK EVALUATION
PSNR does not take aspects of the HVS into consideration although it provides an overall evaluation
We will use another perceptual quality measure called the weighted peak signal to noise ratio (WPSNR)
![Page 35: An Improved Method Of Content Based Image Watermarking Arvind Kumar Parthasarathy and Subhash Kak 黃阡廷 2008/12/3](https://reader036.vdocuments.mx/reader036/viewer/2022062315/5697bf8a1a28abf838c8a8ae/html5/thumbnails/35.jpg)
EXPERIMENTAL RESULTS
Lena reference image Watermarked image WPSNR = 38.99 dB.
Original watermark Recovered watermark
![Page 36: An Improved Method Of Content Based Image Watermarking Arvind Kumar Parthasarathy and Subhash Kak 黃阡廷 2008/12/3](https://reader036.vdocuments.mx/reader036/viewer/2022062315/5697bf8a1a28abf838c8a8ae/html5/thumbnails/36.jpg)
EXPERIMENTAL RESULTS
When we use a slightly bigger watermark of size 15*12 pixels and peak signal to noise value (WPSNR) is again found to be 38.99 dB
Original watermark
Recovered watermark
Watermarked image WPSNR = 35.54
![Page 37: An Improved Method Of Content Based Image Watermarking Arvind Kumar Parthasarathy and Subhash Kak 黃阡廷 2008/12/3](https://reader036.vdocuments.mx/reader036/viewer/2022062315/5697bf8a1a28abf838c8a8ae/html5/thumbnails/37.jpg)
EXPERIMENTAL RESULTS
For this image the scaling factor is 0.084
Original watermark Recovered watermark
![Page 38: An Improved Method Of Content Based Image Watermarking Arvind Kumar Parthasarathy and Subhash Kak 黃阡廷 2008/12/3](https://reader036.vdocuments.mx/reader036/viewer/2022062315/5697bf8a1a28abf838c8a8ae/html5/thumbnails/38.jpg)
EXPERIMENTAL RESULTS
The WPSNR value for the relatively smaller watermarks has been found to be same and it slightly decreases for larger watermarks.
The change in WPSNR values with varying scaling factors
![Page 39: An Improved Method Of Content Based Image Watermarking Arvind Kumar Parthasarathy and Subhash Kak 黃阡廷 2008/12/3](https://reader036.vdocuments.mx/reader036/viewer/2022062315/5697bf8a1a28abf838c8a8ae/html5/thumbnails/39.jpg)
EXPERIMENTAL RESULTS
WPSNR vs. scaling factor plot for Lena and Boat
![Page 40: An Improved Method Of Content Based Image Watermarking Arvind Kumar Parthasarathy and Subhash Kak 黃阡廷 2008/12/3](https://reader036.vdocuments.mx/reader036/viewer/2022062315/5697bf8a1a28abf838c8a8ae/html5/thumbnails/40.jpg)
EXPERIMENTAL RESULTS
Normalized JND values for Lena Normalized JND values for Boat
![Page 41: An Improved Method Of Content Based Image Watermarking Arvind Kumar Parthasarathy and Subhash Kak 黃阡廷 2008/12/3](https://reader036.vdocuments.mx/reader036/viewer/2022062315/5697bf8a1a28abf838c8a8ae/html5/thumbnails/41.jpg)
ATTACKS AND ANALYSIS OF RESULTS
We test the robustness of our scheme for JPEG compression and median filter attack
JPEG compression (q =45)
Recovered watermark
![Page 42: An Improved Method Of Content Based Image Watermarking Arvind Kumar Parthasarathy and Subhash Kak 黃阡廷 2008/12/3](https://reader036.vdocuments.mx/reader036/viewer/2022062315/5697bf8a1a28abf838c8a8ae/html5/thumbnails/42.jpg)
ATTACKS AND ANALYSIS OF RESULTS
JPEG compression (q =40) JPEG compression (q =35)
Recovered watermark Recovered watermark
![Page 43: An Improved Method Of Content Based Image Watermarking Arvind Kumar Parthasarathy and Subhash Kak 黃阡廷 2008/12/3](https://reader036.vdocuments.mx/reader036/viewer/2022062315/5697bf8a1a28abf838c8a8ae/html5/thumbnails/43.jpg)
ATTACKS AND ANALYSIS OF RESULTS
We then test our scheme for its robustness against different types of noise
Uniform Gaussian noise 2%
Recovered watermark
![Page 44: An Improved Method Of Content Based Image Watermarking Arvind Kumar Parthasarathy and Subhash Kak 黃阡廷 2008/12/3](https://reader036.vdocuments.mx/reader036/viewer/2022062315/5697bf8a1a28abf838c8a8ae/html5/thumbnails/44.jpg)
CONCLUSIONS
We employ a better method of detecting edges using phase congruency
The detected corner is used as a factor to establish the uniform regions in the image
![Page 45: An Improved Method Of Content Based Image Watermarking Arvind Kumar Parthasarathy and Subhash Kak 黃阡廷 2008/12/3](https://reader036.vdocuments.mx/reader036/viewer/2022062315/5697bf8a1a28abf838c8a8ae/html5/thumbnails/45.jpg)
CONCLUSIONS
The robustness of our scheme to JPEG compression is found to be very good at a quality factor of 40
A very good balance between robustness and imperceptibility has been achieved using this scheme
![Page 46: An Improved Method Of Content Based Image Watermarking Arvind Kumar Parthasarathy and Subhash Kak 黃阡廷 2008/12/3](https://reader036.vdocuments.mx/reader036/viewer/2022062315/5697bf8a1a28abf838c8a8ae/html5/thumbnails/46.jpg)