Watermarking In The Dct Domain Psychology Essay

Abstract- A blind, low composite, fast, and extremely immune to onslaughts watermarking algorithm is proposed for JPEG compressed images. Our method is a distinct cosine transform ( DCT ) and block based algorithm that adds some Huffman codification words to the terminal of the codification sequence of some blocks during JPEG compaction process. Via a unsighted attack in decipherer, each block is determined to be a watermarked or unwatermarked one. While the hardiness to onslaughts is still preserved, the size of watermarked image may differ from the JPEG compressed one insignificantly. Presented strategy is used for images that have perceptually high quality and due to this, are really disposed to watermarking undertaking for the interest of contents proprietor genuineness. Peak signal to resound ratio ( PSNR ) will be our benchmark to asses the quality of images.

Keywords-watermarking ; image cryptography ; distinct cosine transform ; mean squared mistake

Introduction

Digital image watermarking [ 1 ] , [ 2 ] refers to techniques for implanting extra informations into a host image by deploying the restrictions of human ocular system ( HVS ) . Hidden information can be used as an unseeable label for right of first publication protection, image cleavage, mistake privacy, etc.

Depending on the application, water line informations can be binary sequences ( matching to a consecutive figure or recognition card figure ) , logos, signatures, and etc.

From the position point of water line extraction, there are three types of watermarking strategies consisting Non-blind, blind, and semi blind. Blind methods do non necessitate the original image on which may non be available for the extraction process and rationally a non-blind one needs it. Semi-blind techniques lie merely between these two classs and utilize a key file in order to pull out the concealed information. From another point of view, sometimes watermarking methods are classified to correlativity and noncorrelation based techniques. Correlation based methods normally use direct sequence codification division multiple entree ( DS-CDMA ) techniques in spacial or transform sphere for implanting the water line informations. A more complete

categorization of digital watermarking techniques can be found in [ 1 ] , [ 3 ] .

The chief conflicting demands of image watermarking normally include imperceptibility, hardiness and capacity.

Imperceptibility or transparence represents the invisibleness of the water line embedded in the information. Robustness means that the water line should non be removable by common onslaughts, including filtering, compaction, linear noise, geometric deformations, etc. [ 4 ] , [ 5 ] . Finally, capacity refers to the sum of information ( or warhead ) that can be hidden in the host image and detected faithfully under normal operating conditions [ 6 ] – [ 8 ] .

In this paper we propose a blind and block based watermarking strategy that during JPEG coding procedure conceals crude watermark informations in the DCT sphere. The hardiness of presented method is ensured with the assistance of the fact that we step beyond the traditional tendencies which normally involve in-between and high frequence constituents in the watermarking process and embed the watermark signal in low frequence constituents of DCT sphere.

This paper is organized into four subdivisions. In Section 2 proposed method will be elaborated and in Section 3 simulation consequences will show the range of the hardiness of our method. The decision is provided in Section 4.

PROPOSED METHOD

Watermark Embedding

One of the most popular and comprehensive uninterrupted tone, still frame compaction criterion is the JPEG criterion which is a joint work by the members of the international telecommunication brotherhood ( ITU-T ) and international criterions organisation ( ISO ) . The compaction is performed in four consecutive stairss [ 9 ] consisting degree shifting, DCT calculation, quantisation, and variable-length codification assignment. We note that the coding attack of the most video coding systems ( particularly for I-frames ) is similar two this with normally minor difference. Corollary, this method may be applied to some types of tight picture sequences every bit good.

Presented strategy incorporates DCT calculation and quantisation. For simpleness, we will merely see grey degree images. Approximately talking, during JPEG compaction procedure, at first the image is subdivided into pel blocks of size. Then each block is encountered and its 64 pels are flat shifted by deducting the measure, where is the maximal figure of grey degrees. The 2-D distinct cosine transform ( DCT ) of the block is so computed ( afterlife, we name it matrix ) and quantized in conformity to the undermentioned equation:

( 1 )

is a quantal estimate of and is an component of transform standardization array or so called quantisation tabular array. Obtained matrix is a thin array incorporating a few nonzero coefficients and a big figure of zero-valued coefficients. To group together nonzero coefficients a reordering utilizing a zigzag form is done. While the zig-zag scan of is being done, we reach to a location that has the last nonzero value. We note for future mentions that the brace is the location of the last nonzero constituent in matrix. All the remained constituents of are zero and would be represented or stored via a Huffman codeword matching to the terminal of block ( EOB ) . Now we have reached to the state of affairs that watermark informations may be inserted by the undermentioned manner.

After index, we continue the zigzag scan of matrix and compare the absolute value of with for values that come after during zig-zag scan form. Comparing the absolute value of with, we may make to a location that distance between and is lower than a mention distance ( RD ) as follows:

( 2 )

For some blocks that this state of affairs arises we replace constituent of matrix ( which is unquestionably zero ) with 1 or -1 value if the is a positive or negative value severally. In this manner the Huffman codeword matching to the braces or would be added to the spot sequence of such typical blocks.

The added 1 or -1 are matching to binary figures of 1, 0 or frailty versa. In this manner, a particular binary sequence can be stored in an image by salvaging consecutive spots in the blocks that are prone to watermarking undertaking by the aforementioned manner. For illustration in an image, from top to down and in left to right way, watermarking prone blocks are considered. If the spot that should be saved is indistinguishable to the spot that can be stored in the block, watermarking undertaking is done otherwise the watermarking prone block is discarded and the following suited block is checked.

The added 1 or -1 in the DCT sphere may be unstable for some blocks and removed by onslaughts. Presented method can be refined to be more immune against onslaughts by using the undermentioned facts.

Reinforcing the Added Watermark Signals

Most of the onslaughts target the high and in-between frequence constituents in a given image, because like the watermarking algorithms the onslaught methods should non cut down the perceptual quality of images. Therefore, it is clear that as the location that is set to 1 or -1 goes toward low frequence constituents watermark informations go more stable against such onslaughts.

As the RD value is adopted to be comparatively little, watermark implanting in low DCT frequences will non impact the quality and imperceptibility is preserved. Experimental consequences revealed that the blackened constituents in Fig. 1 ( which are low DCT frequences ) are suited for watermarking by the foregoing scenario. As a affair of fact, afterindex the zigzag scan will go on merely through these constituents. With an appropriate chose of RD, there would be some blocks in the given image that their corresponding is prior to the determined indices.

A quantal DCT block. Blackened squares involve in water line embedding.

As the RD increases the figure of watermarked blocks ( ) increases every bit good, but the added information become more unstable against onslaughts, extraction of water line brushs job, and the PSNR value between JPEG image and its watermarked opposite number beads. Changing RD gives us a freedom to compromise between three demands of watermarking undertaking.

Blind Watermark Detection Algorithm

In a given image ( which is likely a watermarked 1 ) , all blocks that their last nonzero constituent in matrix is 1 or -1 and lies in one of the three foregoing indices are campaigners to be watermarked blocks. This statement is necessary but non adequate status for a block to be a watermarked 1. In fact, in a typical image 1000s of blocks are found that meet this status but are non watermarked blocks. On the other manus, every bit long as the RD is selected low plenty, it is expected that there is no major difference between watermarked and unwatermarked block in pel sphere.

Using these two facts we consider each of the campaigner blocks and replace the last 1 or -1 in matrix with nothing to obtain a new block. To mensurate the difference between the original block and the obtained one in pel sphere, it is proposed to utilize the MSE fiting map that defines between two blocks as follows:

( 3 )

To observe the water line a threshold ( TH ) is adopted. If the is lower than or equal to TH, water line is detected otherwise, the campaigner block is non a watermarked 1. In order to decline false dismaies, TH should be sufficiently little. TH is delivered to the 1s that have permission for water line extraction.

Simulation consequences in the following subdivision corroborate a faithful extraction by utilizing this method.

Simulation RESULTS

Watermark Embedding and Extraction

Fig. 2 indicates two watermarked images with RD=9.5, their corresponding JPEG images, the difference image between them and the extracted water line.

( a ) JPEG Images

( B ) Watermarked Images

( degree Celsius ) Difference Images

( vitamin D ) Extracted Watermark

Top to toss off: JPEG images, watermarked images, extremely amplified difference images between watermarked and unwatermarked images, and extracted water line ; left to compensate: Lena, Pirate.

Difference images between watermarked JPEG images and strictly JPEG images ( with a 50 % quality factor ) have been extremely amplified to do the watermarked blocks rather seeable. Images are perceptually identical. High PSNRs corroborate this fact. The PSNR value between watermarked and JPEG Lena image is 52.37 dubnium with N=45 watermarked blocks and by analogy for the Pirate image has been evaluated to be 53.17 dubnium with N=31. It is besides seen that extracted water line and difference images are matched together which in bend corroborates precise sensing of water line with no false dismay. In extracted water line images, rather white blocks are related to the blocks that 1 is set alternatively of nothing and darker blocks show the 1s that -1 is replaced. In all experimental consequences that are provided in subdivision 3 the TH is by experimentation adopted to be 3.9. For Lena and Pirate images it is the minimal value that the extraction is done successfully. It is apparent that by minimising the TH, hardiness versus attacks additions.

Robustness to Attack

In this subdivision, hardiness of presented method is inspected. For the first onslaught type, linear white Gaussian ( AWG ) noise and salt & A ; pepper ( S & A ; P ) noise are considered and simulation consequences for Pirate image are summarized in table 1. It is seen that as the denseness of noise is low plenty so that the noise add-on is visually unobtrusive, watermark information is extracted successfully. Meanwhile, by increasing the noise denseness, the PSNR beads, false dismaies emerge and complete extraction fails. In these instances, resulted image is a extremely reduced quality and worthless one.

In fact, meanwhile the added 1 or -1 is still preserved, noise add-on with a comparatively high denseness adds some high frequence constituents to matrix and a watermarked block is non even a campaigner block for water line sensing algorithm.

Table I

Extraction per centum for different RDs and noise densenesss.

Noise Type

RD

Extraction ratio

PSNR ( dubnium )

AWG

8.2

100 %

100 %

failed

53.9

48.4

29.9

AWG

9.0

100 %

50 %

failed

53.4

47.9

29.9

S & A ; P

8.2

100 %

100 %

100 %

53.9

52.3

35.6

S & A ; P

9.0

100 %

100 %

91 %

53.4

51.0

35.3

.

Another common onslaught type is additive and nonlinear 2-D filters that may be exerted to images. Some of these spacial filters do non slake the PSNR perceptibly and have no obvious consequence on the ocular visual aspect of image. Such filters may see as an onslaught to watermarked image. In table 2 simulation consequences are tabulated for two common and good known filters. As table 2 shows, taking RD value sufficiently little assures a good hardiness to filtrating onslaughts that is rather imaginable, because water line informations is more stable for little values of RD.

Listed PSNR values are the PSNRs between watermarked and filtered images versus their lone watermarked opposite numbers which are comparatively fixed with RD alterations. Robustness versus other filter types can be found in table 4.

TABLE II

Common filtrating onslaughts to Lena image with different RDs.

Filter type

Extraction per centum

PSNR

( dubnium )

RD=8.21

RD=8.22

RD=8.23

RD=8.3

Gaussian

100 %

100 %

90 %

90 %

43.0

Median

100 %

100 %

60 %

50 %

37.4

One of the most sophisticated onslaught types is the rotary motion of image with a little angle. Almost all watermarking algorithms that their water line extraction depends on the precise location of watermark signal, suffers from this sort of onslaught.

To pull out the water line signal, with the assistance of existed methods [ 10 ] the angle of rotary motion is estimated and, if possible, watermark signals are extracted. Our method for gauging the angle of rotary motion is a grid base one that can be found in [ 10 ] . In [ 10 ] a method is proposed to add a grid to an image that can be used to scale, revolve, and switch an image back to its original size and orientation. The grid is represented by a amount of sinusoidal signals, which appear as extremums in the FFT frequence sphere. These extremums are used to find the geometrical deformation. This grid is added to image except for the watermarked blocks.

Fig. 3 indicates a revolved version of watermarked image and the water line informations that are extracted from its rotary motion compensated original opposite number. While the appraisal undertaking is done accurately the complete extraction of water line is guaranteed with an reverse rotary motion and using watermark sensing algorithm.

Left to compensate: rotated version of watermarked image with RD=9 and angle of rotary motion equal to, extracted water line from rotary motion compensated version.

Extinguishing a row or column of a given image or cropping is another type of geometrical onslaught. Equally long as the eliminated row or column is interpolated via next pels and the image dimensions remain integral, watermark extraction encounters no trouble due to the correlativity between neighbouring pels. In Fig. 4 eight indiscriminately selected column of Lena image was eliminated and as an alternate the nearest left column was repeated. Comparing the extracted water line with the 1 in Fig. 2, it is seen that all watermarked blocks are detected right.

In this experiment five Numberss of eliminated rows have selected to stop some watermarked blocks.

Left to compensate: eight indiscriminately selected columns are white and distinguishable, extracted water line.

Comparison with SIMILAR METHODS.

In [ 11 ] a CDMA based watermarking strategy is presented which works in DCT sphere and offers a good hardiness versus onslaughts, particularly for filtrating type. In this method a pseudo random form is added to middle and low frequence ( except DC constituent ) constituents in DCT sphere. More inside informations may be found in [ 11 ] but implanting algorithm may be described mathematically through the equation below.

In the above equation, represents DCT constituents of each block and is a pseudo random form in DCT sphere. represents in-between and low frequences of DCT sphere and K is a addition factor. Detection of water line is correlativity based [ 11 ] .

In [ 12 ] besides a DCT based method is presented which dainties low frequence of DCT sphere. In this method, five low frequence constituents ( except DCs one ) are estimated via DC constituents of four adjacent blocks. Writers do non claim that their appraisal equations are optimal but experimental consequences are quite satisfactory.

Here during two tabular arraies these three methods are compared versus JPEG compaction with different quality factors and filtrating onslaughts. In both experiments, Lena image is watermarked with N=22 by all the three methods to convey about equal warhead status. Other parametric quantities are selected to be RD=8.21 and k=3.

In table 3, proposed method is compared with the two other similar strategies from the position point of hardiness versus JPEG compaction with different quality factors. As the corresponding spot error rates ( BER ) show, our method either is more robust than or every bit robust as the two other blind strategies.

All the PSNR values are measured in dB graduated table. For our method they are measured between the watermarked image with the standard quantisation tabular array ( which leads to a 50 % quality factor ) and the 1s which yield other quality factors. For CDMA based and appraisal based methods the PSNRs are measured between the original watermarked image and its JPEG coded opposite number.

Table Three

This tabular array shows a comparing between the three watermarking methods versus JPEG compaction with different quality factors.

JPEG quality factor

Our method

CDMA based method

Appraisal based method

BER

PSNR

BER

PSNR

BER

PSNR

60 %

0 %

50.01

0 %

46.50

0 %

43.05

50 %

0 %

47.36

0 %

45.03

5 %

39.69

45 %

0 %

39.83

4.43 %

36.60

5 %

37.26

40 %

22.7 %

38.40

10.12 %

33.35

8.4 %

35.05

30 %

37.3 %

30.50

30.15 %

30.54

50 %

31.92

In table 4, three aforesaid methods are tasted against filtrating onslaughts. Here we consider four types of filters. As the corresponding spot error rates demonstrate, the public presentation of proposed method is rather competitory in comparing with the two other methods. In fact, CDMA based methods normally are vulnerable to onslaughts that blur images. Because low frequence contents interfere with the water line and the correlativity based sensor fails to pull out the water line right. Due to this, normally a matched filtering before watermark sensing improves watermark sensing significantly [ 13 ] . This decreases the part of the original image to the correlativity.

Table Four

This tabular array shows a comparing between the three watermarking methods versus different filtrating types.

Filter type

Our method

CDMA based method

Appraisal based method

BER

PSNR

BER

PSNR

BER

PSNR

Gaussian

0 %

43.0

5 %

37.2

25.3 %

35.5

Averaging

0 %

30.4

6.4 %

32.4

33.3 %

33.4

median

0 %

37.4

0 %

34.0

15.4 %

34.2

Wiener

0 %

31.8

0 %

29.5

0 %

31.7

decision

Presented method in malice of its simpleness has a satisfactory hardiness to common types of onslaughts. It shows the capacity of DCT coefficients for water line embedding and besides reveals that water line implanting in low DCT frequences is besides possible. Proposed method offers an RD parametric quantity which by changing it a via media between transparence, hardiness and capacity use is possible and the size of watermarked image is governable.