Image Watermarking with DCT-Optimization based Psychovisual Threshold and Fuzzy Adaptive Median Filtering (FAMF) for Noise Distortion

: The high speed development of internet technology, illegal copy, transmission and distribution of digital multimedia has emerged to be a significant security challenge. This challenge inspires the design of a solution for image authentication and copyright protection. Recently, an optimal Discrete Cosine Transform (DCT) psychovisual threshold is proposed for digital watermarking approach. The newly introduced approach also yields a distorted watermark extraction with image-rotation attack. This problem will be focused in this research work. In order to resolve this issue, Fuzzy Adaptive Median Filtering (FAMF) is introduced for noise elimination in the distorted watermark extraction suffering from image-rotation attack. In the newly introduced work, embedding regions are decided on the basis of the lowest modified entropy value of the image blocks. In this, the optimization of the lowest modified entropy value is done by using Bat Algorithm (BA). Therefore, the optimal psychovisual threshold is decided for embedding the watermark in the host image for getting a better image quality. The lowest modified entropy value specified the greatest redundant image information. The scrambling of the watermark bits are done prior to them being embedded into the chosen coefficients. The newly introduced approach has been tested under various kinds of attacks. The newly introduced approach also exhibits the superior performance in terms of reliability under several numbers of combined attacks.


INTRODUCTION
The high speed development of internet technology, illegal copy, transmission and distribution of digital multimedia has emerged to be a significant security challenge.This challenge inspires the design of a solution for image authentication and copyright protection.Digital watermarking is regarded to be as an alternate solution to stop illegal copy and it has gained extensive focus (Ernawan et al 2016) (Kumar et al 2017).The watermark insertion and extraction may be performed by the owner to guarantee and verify its ownership and authenticity by making use of theprocess of digital watermarking.Generally, watermarking can be carried out either in spatial domain in which the embedding of the watermarks are done in the image pixels directly or in the frequency domain where the insertion of the watermark is inserted in the frequencies acquired through frequency transformation of the image (Moosazadeh and Ekbatanifard, 2017).(Kumar et al 2017).Watermarking approaches dependent on frequency domains with the transformations like Discrete Cosine Transform (DCT), Discrete Wavelet Transform (DWT) and Integer Discrete Wavelet Transform (IDWT) have been employed in the past decade owing to greater reliability and imperceptibility.But, DWT and IDWT are affected by huge computational complexity.
A robust digital watermarking approaches to meet certain characteristics (Chetan and Nirmala 2015) (Hsu et al 2015).Blind Image Watermarking through Exploitation of Inter-Block Prediction And Visibility Threshold in DCT Domain like Imperceptibility, Robustness, Undetectable, Security And Blind Extraction.The embedded watermark into the host image has to generate less distortion so that the detection of the watermark effect cannot be done by the human visual system.
The watermarking methods have to guarantee that the quality is close to the actual host images.Watermarks needs to tolerate the common signal processing and geometrical attacks e.g., filtering, noise addition, image compression, size change, cropping and change of pixel values.Owing to the restricted size of the watermark, which can be embedded into a host image, a binary watermark remains a desirable selection.
The objective of a watermarking approach is to protect the watermark present inside the image so that the watermark is very hard to be identified and corrupted by attackers.A blind watermarking approach yields an independent recovery of the watermark with no reference to the actual host image.It is a huge challenge to design a blind watermarking approach, which is reliable and imperceptible with a good security characteristic and low computational complexity.Abdul et al (2013) developed the watermarking algorithms to be reliable against deliberateor unintentional attacks like JPEG compression, additive white Gaussian noise, low pass filter and color attacks employing error correcting codes.The usage of error correcting codes in an addition mode lets the non-binary block codes to exhibit good performance against JPEG compression, noise and brightness attacks.

LITERATURE REVIEW
Ernawan et al (2014) suggested a high compression approach, which depends on psychovisual threshold in the form of an adaptive image compression.The lower bit rates of image compression employing scaling factor will deteriorate the visual quality on the images that are undergoing compression.The comparison between JPEG-3 image compression employing the common quality factor and an adaptive psychovisual threshold has been performed.The adaptive quantization tables dependent on psychovisual threshold exhibit an enhancement on the quality of image reconstruction at the lower average bit length of Huffman code.
Jeswani and Sarode (2014) studied about a new blind watermarking approach for color images with DCT.In the newly introduced approach, the cover image is divided into 3 planes, which are R, G, B and for watermarking purposes, B plane is chosen.B plane is segmented into blocks with size 8×8.DCT is used on every 8×8 sized block and the middle frequency coefficients are chosen for embedding the watermark.At last, the binary watermark is embedded into every 8×8 DCT block by modifying middle frequency coefficients DCT (4, 3) and DCT (5,2).The newly introduced approach does not need cover image during the time of watermark extraction and then also the watermark is extracted wholly and offers better performance compared to its counterparts.Abu and Ernawan (2015) introduced a psychovisual threshold on the huge DCT image block that will be utilized for automatic generation of the much required quantization tables.The psychovisual threshold denotes the sensitivity of the human visual perception at every frequency order to the reconstruction of the image.One ideal contribution made by the transform at every frequency order will become the primitive of the psychovisual threshold in image compression.The newly introduced psychovisual threshold will be utilized for prescribing the quantization values at every frequency order.The psychovisual threshold on the huge image block yields considerable enhancement over the output images' quality.
Ernawan (2016) demonstrated a watermark embedding approach that depends on the psychovisual threshold and edge entropy.The embedding of a natural watermark that relies on the characteristics of the human eye can be used for efficiently hiding a watermark image.The sensitivity of slight changes in DCT coefficients against JPEG quantization tables was examined.A watermark embedding approach was developed, which provides good resilience against JPEG image compression.The newly introduced approach was tested under various kinds of attacks.The newly introduced approach can attain a higher imperceptibility and reliability against attacks.The process of watermark recovery also offers robustness against attacks.
Ernawan et al (2017) examined an enhanced watermarking that depends on 4×4 DCT-Singular Value Decomposition (SVD) blocks employing modified entropy in image watermarking.The modified entropy is utilized for selecting imperceptible blocks.The novel watermarking approach uses the lowest entropy values to decide the imperceptible areas of the watermarked image.The newly introduced watermarking approach generates a superior level of reliability and imperceptibility of the watermarked image against various attacks.The proposed approach exhibits the enhancement with regards to structural similarity index and normalized correlation of the watermarked image.
Moosazadeh and Ekbatanifard (2017) introduced a digital image watermarking algorithm in YCoCg-R color space.For embedding the watermark, the newly introduced technique makes use of Discrete Cosine Transform and its coefficients association.During the process of watermark bits embedding, the degree of complexity of host image blocks is used for selecting the target blocks and the energy value of host image blocks is utilized for adaptively selecting the embedding strengths, which results in an extremely high robustness, particularly against JPEG compression.The Arnold transformation, also, has a significant part in improving the security of the proposed technique through the scrambling of the watermark.Taking the conflict among three important needs of watermarking into consideration, inclusive of imperceptibility, robustness and capacity, the comparison of the proposed technique is done with other algorithms and the results indicate the improved robustness of the newly introduced technique compared to the other algorithms having the same capability.
Roy and Pal (2017) suggested color multiple watermarking technique that depends on DCT and repetition code.At first, green and blue components of color host image are chosen for the insertion of several watermarks.Thereafter, every green and blue component of the image is divided into non overlapping blocks and next, DCT is used on every block.In this method, a binary bit of watermark is embedded into transformed block of the green/blue component by changing few middle important AC coefficients employing repetition code.During the embedding of multiple watermarks in green and blue components of the newly introduced technique, DC and few higher AC coefficients are maintained intact after the zigzag scanning of every DCT block to guarantee the imperceptibility of the watermarked host image.

PROPOSED METHODOLOGY
Fuzzy Adaptive Median Filtering (FAMF) is proposed for noise reduction in the distorted watermark extraction under image-rotation attack.The proposed work, embedding regions are determined based on the lowest modified entropy value of the image blocks.Here the lowest modified entropy value is optimized via the use of Bat Algorithm (BA).Thus, the optimal psychovisual threshold is determined to embed the watermark in the host image for the best image quality.The lowest modified entropy value indicates the highest redundant image information.The watermark bits are scrambled before they are embedded into the selected coefficients.Embedding regions are determined based on the lowest modified entropy value of the image blocks.The lowest modified entropy value indicates the highest redundant image information.Binary watermark bits are scrambled before they are embedded into the selected coefficients.The proposed technique has been tested under different types of attacks.Test results have been verified with other schemes in terms of Normalized cross-Correlation (NC) and Structural Similarity (SSIM) index (See figure 1).

Fuzzy Adaptive Median Filtering (FAMF) (a) Noise model
In our Noise Model, pixels are randomly corrupted by some extreme values, generated with the equal or unequal probabilities.A random valued impulse noise model with error probability p , input intensity value I (x, y) and original intensity value O(x, y)can be represented as follows: Where S(x, y) and P(x, y) are -salt‖ noise and -pepper‖ noise with probability and , p = + is the noise density.

(b) Estimation method
The proposed algorithm is applied to each pixel of the noisy image in order to identify the level of the noise contamination.In this section, we explain a new method to estimate whether the image is corrupted by strongly impulsive noise.The steps of noise degree estimate are described as follows.

Figure 2. Neighborhood of a central pixel (x, y)
For each of the four neighbors of I (x, y) the gradient values are shown in the table 1.Each Direction (column 1) corresponds to a position (Figure 2).Column 2 represents the gradient for each direction, column 3 lists the related pixels for calculating the gradient value.As is known to all, image pixels that are corrupted with impulse noise will generally have large Grad(x, y) values, because these pixels mostly occur as outliners in comparison to their neighborhood.However, edge like regions also have large Grad(x, y) values.Therefore we use Smax (x, y) and Smin (x, y) to solve this problem.Then the sum of gradients and the two sums of extreme values are connected together into one single value called -fuzzy corrupted value‖.This fuzzy corrupted value is then used to determine if filtering window is highly corrupted by impulse noise or not.This can be implemented by the following Fuzzy Rule.Fuzzy Rule: Defining the fuzzy corrupted value .IF Grad(x, y) is large AND Smax (x, y) is larger OR Grad(x, y) is large AND Smin (x, y) is larger THEN λ c is large.THEN the pixels in a certain filtering window are highly corrupted by impulse noises.

Figure 3. The membership function LARGE
In this rule, larger means that most of the pixel values are equal to the extreme values in filtering window.Large can be represented as a fuzzy set (Ernawan et al 2017).A fuzzy set in turn can be represented by a membership function.An example of a membership function LARGE (for the fuzzy set large), is pictured in Figure 3.If the Grad(x, y) for example has a membership degree one (zero) in the fuzzy set large, it means that the level of noise contamination in this filtering window is considered as (not) high for sure.Membership degrees between zero and one indicate that the Grad(x, y) value is large in a certain degree.For more background information about fuzzy logic we refer to C. Lai, An Improved SVD-Based Watermarking Scheme Using Human Visual Characteristics [40].The parameters a and b of Figure 3 can be defined as follows: ( ) Of course, a is not always equal to 0, it can be a larger value than 0, for example, any values in the range of [0,9].
Here n is the actual number of pixels in filtering window.Commonly, n is equal to (2K +1)× (2K +1), but if window is located at the edge of the image, n will be smaller than (2K +1)× (2K +1).Experimental results have shown that the best choice for parameter k is k = 168 .frequency modulated sound pulses or constant frequency sound pulses.This echolocation behavior of bats is devised to optimize a given objective function ( Yang 2010).The starting point in BAT based optimization algorithm is to define a objective function f(x).Initialize a few parameters such as bat population (i=1 to n ) and initial velocities of bats .Initialize pulse frequency .Pulse frequency can vary in the range , -= 1 to n .For a particular optimization function the solution space can be adjusted by selecting frequency range close to sphere of interest.Finally initialize pulse rate and loudness By selecting iterations itr to a specific count, new solutions are generated by adjusting the frequencies.This means the bats try to reach their target location by adjusting their frequencies and computing their velocities and locations at each new frequency.This computation can be mathematically modeled as Where µ [0,1]is a uniformly distributed random vector.
Where is the initial velocity of the bat population.*x is the current global best solution ,i.e.location of the target which is located after comparing all the solutions among all the n bats.The final or next best location of the target can be updated using the following mathematical mode From equation (11) it can be seen that the location or solution space is updated with the bat updated velocity in each iteration in turn is derived from bat frequencies or wavelengths as or Initially bat frequencies are randomly allocated between [ which can be chosen based on the objective function that is being minimized.The above process provides a solution space containing global best positions of the ‗n' bats target.A local best solution can be generated locally for each bat using random walk = ( 13) is the new location or solution space that is produced from , the old global solution space using a loudness updation factor with a random constant η − [ 1,1] Further updation of loudness and pulse rate are done between iterations.initial loudness can be set as 1 and the final loudness that is to be reached is set as 0 indicating that the bat has finally reached its target where it stops temporarily its search process.The new loudness factor can be mathematically modeled as ( 14) Where κ is a constant.The pulse rate is updated as Where ρ is a constant.For experimentation both the defined constants are given values ranging from 0 to 0.7 based on various medical cover images used for watermarking.A series of experiments are to be conducted to fix the values of loudness and pulse rate .For watermarking simulations using medical images as cover images the value of loudness varies , pulse rate , -.The selected coefficients are ordered into a vector of coefficients as shown in Figure 4.4.Also assume that those locations provide less distortion and more robustness under image compression attacks.
In the algorithm, T represents a threshold value obtained from the trade-off between the imperceptibility and robustness of the watermarked image under JPEG compression.T is measured based on the relationship between SSIM and NC values.Embedding blocks are selected based on the modified entropy values.The number of selected blocks is considered to be the same as the number of watermark pixels.
In this research, use a binary watermark with the size 32*32 pixels.Therefore, 1024 of 4096 blocks are selected to embed the watermark logo.In the watermarks are embedded in the frequency coefficients of each selected block of an image using the technique given in step 8.Note that for x = 0, 1, and 2, A(2x) represents A (0), A(2), A (4) and A(2x + 1) denotes A (1), A(3), A(5), respectively as shown in Figure 4.4 and present variant thresholds for watermark embedding.If A(y) < 0, for y =0,1,....5, the threshold value is negative; otherwise, the threshold is positive.The locations of watermark embedding are stored to determine the selected blocks during watermark extraction.Selected blocks based on modified entropy are chosen which has large redundant data.Attackers may identify the selected block, while they will find it difficult to identify the scrambled watermark.

Step 1: Start
Step 2: Read the water marked input image Step 3: The water marked image is divided into 8*8 pixels non overlapping block sing Fuzzy Adaptive Median Filtering (FAMF).
Step 4: Compute modified entropy by Bat Algorithm(BA) for each non overlapping blocks.
Step 5: Compute the lowest modified entropy values and save their x and y coordinates of selected blocks.
Step 6: The binary watermark is scrambled using Arnold chaotic map.
Step 7: Selected binary watermark blocks are converted into frequencies using DCT.
Step 8: DCT coefficients are converted into a vector by using zig-zag order.
Step 9: Obtain the modify coefficient pairs by compaining the binary watermarked image and the zig-zag order select locations.
Step 10: The scrambled binary watermark is inversed by Arnold chaotic map to obtain the original watermark.
Step 11: Finally got the original water marked image.

RESULTS AND DISCUSSION
In this system, a binary logo image with 32*32 pixels is used as a watermark.The logo is scrambled by Arnold chaotic Encryption Image for additional security.In this section deals with comparision analysis between optimal discrete cosine transform (DCT) psychovisual threshold and Fuzzy Adaptive Median Filtering (FAMF).The proposed system simulated using MATLAB.The input image of existing system is shown in figure 5.The watermarking input image is shown in figure 6.The Arnold chaotic Encryption Image is shown in figure 7. The figure 8 shows the water marked image of optimal discrete cosine transform (DCT) psychovisual threshold with the distorted watermark extraction.The figure 9 shows the Arnold chaotic Encryption Image of proposed system.The figure 10 shows the watermark Extracted Image of proposed system.The figure 11 shows the Watermarked Image.The figure 12 shows the Rotated Watermarked Image.The figure 13 shows the watermark Extracted Image of the proposed system.In this proposed system, with the help of Bat algorithm the embedding regions are determined based on the lowest modified entropy value of the image blocks using a new Fuzzy Adaptive Median Filtering (FAMF) algorithm to reduce the distortion in the watermarking.The mean squared error (MSE) or mean squared deviation (MSD) of an estimator (of a procedure for estimating an unobserved quantity) measures the average of the squares of the errors-that is, the average squared difference between the estimated values and what is estimated.MSE is a risk function, corresponding to the expected value of the squared error loss.The fact that MSE is almost always strictly positive (and not zero) is because of randomness or because the estimator does not account for information that could produce a more accurate estimate.Mean Squared Error defined as

PEAK SIGNAL-TO-NOISE RATIO, (PSNR ) and ABSOLUTE RECONSTRUCTION ERROR (ARE)
A higher PSNR generally indicates that the reconstruction is of higher quality, in some cases it may not.One has to be extremely careful with the range of validity of this metric (Huynh-Thu (2008) and Huynh-Thu et al (2012).Generally, PSNR has been shown to measure the quality of images by humans defined by Imperceptibility is measured by Absolute Reconstruction Error (ARE) defined by

STRUCTURAL SIMILARITY (SSIM)
SSIM is used for measuring the similarity between two images.The SSIM index is a full reference metric; in other words, the measurement or prediction of image quality is based on an initial uncompressed or distortion-free image as reference.SSIM is designed to improve on traditional methods such as peak signal-to-noise ratio (PSNR) and mean squared error (MSE).

( ) , ( )-[ (( ))] ., ( )-(18)
Bright field images between scans (normal scan and dye switch scan) were aligned by computing the normalized cross-correlation matrix of pairs of images in MATLAB.A "globally optimized," normalized crosscorrelation algorithm is used to achieve sub-pixel alignment accuracy is defined by, A figure 14 shows the comparison results of the PSNR using proposed and existing method.The PSNR comparison results are shown in y-axis and methods are represented in the X-axis.It concludes that the proposed FAMF -BA algorithms produces higher PSNR results of 41.09% which is 5%, higher when compared to existing DCT methods respectively.

Figure 15. Comparison results of the Water marking MSE using proposed and existing method
A figure 15 shows the comparison results of the MSE using proposed and existing method.The MSE comparison results are shown in y-axis and methods are represented in the X-axis.It concludes that the proposed FAMF -BA algorithms produces lesser MSE results of 0.0198 % which is 0.019%, lesser when compared to existing DCT methods respectively.

Figure 16. Comparison results of the Water marking ARE using proposed and existing method
A figure 16 shows the comparison results of the ARE using proposed and existing method.ARE comparison results are shown in y-axis and methods are represented in the X-axis.It concludes that the proposed FAMF -BA algorithms produces higher ARE results of 0.013 % which is 0.005%, higher when compared to existing DCT methods respectively.

Figure 17. Comparison results of the Water marking SSIM using proposed and existing method
A figure 17 shows the comparison results of the SSIM using proposed and existing method.The SSIM comparison results are shown in y-axis and methods are represented in the X-axis.It concludes that the proposed FAMF -BA algorithms produces higher SSIM results of 0.98% which is 0.1%, higher when compared to existing DCT methods respectively.

Figure 18. Comparison results of the Water marking NXC using proposed and existing method
A figure 18 shows the comparison results of the NXC using proposed and existing method.The SSIM comparison results are shown in y-axis and methods are represented in the X-axis.It concludes that the proposed FAMF -BAT algorithms produces higher NXC results of 0.9989% which is 0.1%, higher when compared to existing DCT methods respectively.

CONCLUSION AND FUTURE WORK
In this work one new method is proposed for noise reduction in the distorted watermark extraction.Here the lowest modified entropy value is optimized via the use of Bat Algorithm (BA) in watermarked extraction process and using BAT algorithm to optimize the values of PSNR and NCC to produce a good embedding strength for the watermark.The proposed technique has been tested under different types of attacks.In this proposed technique, efficient water mark could take place using Fuzzy Adaptive Median Filtering (FAMF).At last performance is analyzed by means of using various lists of parameters such as recall, precision, f-measure and finally accuracy, SSIM, NXC.In future work, this automatic watermark extraction technique has been enhanced by adopting various other unsupervised learning algorithms and techniques might be use and performance would increase.
First for a pre-defined (2K +1)× (2K +1) (K = 1,2,........, N) filtering window around a certain image pixel I (x, y) at position (x, y) we account the number of the max and min values in filtering window: Start Read the original image Divide 8*8 pixels of non overlapping blocks( famf) Compute the modified entropy Compute lowest entropy values by selected 8*8 blocks Save all the x and y coordinates of selected blocks DCT Zig zag order to select locations Modify coThe SIJ Transactions on Computer Science Engineering & its Applications (CSEA), Vol. 6, No. 5, October 2018 ISSN: 2321-2381 © 2018 | Published by The Standard International Journals (The SIJ) 40 Second for each pixel in a certain (2K +1)× (2K +1) filtering window around I (x, y) calculate its four gradient values.Here we use a 3×3 neighborhood window for example as illustrated in Figure 4.1.The four neighbor of (x, y) corresponds to the direction: {E = East, SW = SouthWest , S = South, SE = SouthEast } .Each such direction with respect to (x, y) corresponds to a certain position.

Table 1 .
Gradient values to calculate the level of noise contamination.Here define the gradient value as: Then calculate the sum of four gradient values of pixel (x, y) and at last get the sum of four gradient values of each pixel in filtering window.we calculate the sum of four gradient values of pixel (x,y) and at last get the sum of four gradient values of each pixel in filtering window.