Documents

A DWT-based Image Watermarking Approach Using Quantization on Filtered Blocks

Description
Wavelet's application
Categories
Published
of 6
All materials on our website are shared by users. If you have any questions about copyright issues, please report us to resolve them. We are always happy to assist you.
Related Documents
Share
Transcript
  A DWT-based Image Watermarking Approach usingQuantization on Filtered Blocks Thien Huynh-The  ∗ , Sungyoung Lee  ∗ , Hieu Pham-Chi  † , Thuong Le-Tien  †∗ Department of Computer EngineeringKyung Hee UniversityGyeonggi-do, 446-701, KoreaEmail: thienht@oslab.khu.ac.kr; sylee@oslab.khu.ac.kr † Department of Electric and Electronics EngineeringHochiminh University of TechnologyHochiminh City, VietnamEmail: phamchihieu3006@gmail.com; thuongle@hcmut.edu.vn  Abstract —In this paper, the authors propose a digital water-marking scheme for gray-scale images, in which, coefficients of a host image are quantized in the wavelet domain. In particular,the binary watermark image is embedded into the wavelet-blocksthat can be obtained by grouping four coefficients at differentsub-bands at corresponding coordinates. The modification of onlytwo coefficients at the LH4 and HL4 sub-bands is decided by thevalue of watermark bit and the significant difference of them.In order to improve the performance, these blocks are filteredby sorting in the difference as the reparation before applying anembedding algorithm. The improvement reduces the degradationin the output image. The method is evaluated for many samplesunder various types of attack including the geometric and non-geometric attacks. Moreover, the comparison of the proposedmethod and recent others is also represented at the same payload. Keywords - Within-Class Variance; Wavelet-Block; Coefficient Quantization; Blind Watermarking; Otsu Method; Filtered Blocks. I. I NTRODUCTION The rapid growth of the Internet have given customermore changes and challenges in the transmission, storage, andsharing of data. Therefore, the customer need to face and dealmany troubles about the preservation of unauthorized data.Hence, the digital watermarking technique is suggested andapplied to multimedia products to solve issues related to thecopyright protection.The watermarking system can be examined as the non-blind, semi-blind, and blind scheme according to whetherthe srcinal data in the extraction process. In additions, theyare also classified based on the transformation domains suchas, Discrete Wavelet Transform [1]-[2]-[3], Discrete FourierTransform [4], and Curvelet Transform [5]. Bhatnagar et al.[1] have proposed the method in which the host image is de-composed by the fractional wavelet packet transform to embeda watermark into the non-overlapping blocks. By modifyingthe block singular values corresponding to the watermarkingsingular values, the method reduced the distortion in outputimages. With the combination of an image secret sharing, JustNoticeable Distortion (JND), and wavelet transform, Hsieh etal [2] have recommended the two-phases method for colorimages. Another approach extending JND to Just NoticeableColor Difference (JNCD) for color images was recommendedby Chou et al [6]. Using a combination of the Discrete WaveletTransform and the Singular Value Decomposition (SVD),Song et al [7] proposed a robust watermarking technique basedon the different regions of the host images. The performanceof a watermarking process can be improved by utilizing thewatermark data which frequency spectrum not dissimilar tothat of the host. Wang et al [4] have recommended a novelsolution based on quaternion Fourier transform and LeastSquares Support Vector Machine (LS-SVM). According tothe modulation of Fourier coefficients, the binary watermark is embedded into srcinal color images. In order to extractthese bits, the LS-SVM correction is utilized. The Curvelettransform has employed in [5] as the decomposition toolto change the domain in both embedding and extractingprocess. In particular, Song et al decided to choose the middle-frequency sub-bands of images to employ a watermarkingalgorithm to achieve a robustness. However, the method cannotbe strong under hue-based attacks and also the compression.Some existing approaches have proposed to only againstgeometric attacks of rotation, scaling, and cropping [8]-[9].These method usually do not use techniques of the trans-formation domain. As in [8], Tuan et al have proposed afast algorithm to exploit the correlation among all of thedetected regions for an embedding process based on theScale Invariant Feature Transform (SIFT) and the even-oddquantization technique. In order to increase the security, asecret key associated with the SIFT factors is suggested todistribute regions randomly. In [9], Ping et al simultaneouslyproposed two approaches to be robust under geometric dis-tortions. With the first approach, the authors proposed a newnormalization procedure which is invariant to affine transformattacks. In the second approach, a resynchronization schemeis recommended to alleviate the effects of random bendingattacks. To be contrast together, one is suitable for publicwatermarking applications and another is good for privateapplications. However, as a novelty, the direct-sequence CodeDivision Multiple Access technique (CDMA) is employed inboth schemes. Applying machine learning into a watermarking  LL4 LH4 HL4 HH4 LL4 LH4 HL4 HH4HL3 HH3LH3 HL2LH2HH2 HL1 HH1LH1 (1,1) (1,1) (1,1) (1,1)(1,2) (1,2) (1,2) (1,2) 1 st  BLOCK Fig. 1. Grouping the coefficients into blocks. framework is also recommended by Fu et al [10]. In the detail,the Support Vector Machine (SVM) is trained to extract thewatermark. Although the information is successfully extracted,the robustness is limited by geometric attacks. In recent years,the Genetic Algorithm - Back Propagation Network [11] isalso employed for an extraction stage based on the DWT orDCT domain by Agarwal et al.In this paper, the authors also transform the host image intothe wavelet domain to employ the embedding or extractingalgorithm. In stead of establishing a block of one coefficientsat the high-level sub-band and four ones at the lower sub-bandin [3], a block in proposed method includes four coefficients of sub-bands at the same coordinate. These blocks are then usedto embed watermark bits by altering two coefficients at middle-frequency sub-bands. In order to minimize the distortion of output images, wavelet blocks will be sorted and filtered tobe appropriated with watermark bit values. To recover thewatermark image, the different values of modified blocks arecompared with an adaptive threshold which is determinedbased on the Otsu method. As an organization of this paper,the section II represents the main method in details. Theexperimental results and conclusion of this work are thendescribed in section III and IV, respectively.II. T HE PROPOSED METHOD In this research, a  32 × 32  binary watermark image is hiddeninto  512  ×  512  gray-scale images by applying the 4-levelDWT technique for cover images. In particular, watermark bits are embedded into wavelet-blocks which are generatedby grouping each coefficient of 4-level sub-bands at the samecoordinate. To be clear, each block contains four coefficients { LL 4 i ,HL 4 i ,LH  4 i ,HH  4 i }  for  i th coordinate as shown inthe Fig. 1. With this way, there are  32  ×  32  blocks corre-sponding to number of embedded bits. The different values be-tween two coefficients of middle-frequency sub-bands, calledsignificant difference, are then computed for all of blocks. TheLL and HH coefficients are not chosen for an implementationbecause the low-frequencycoefficients are sensitive with noise,meanwhile, the high-frequency coefficients are easy to beeliminated by the JPEG compression. Thus, the LH and HLcoefficients will be used to calculate different values throughthe following equation: d i  =  | LH  i  − HL i |  (1)where  d i  is the difference of   i th block. The Fig. 2(a) showsthe number of blocks in different value of a sample image.  A. The embedding algorithm In this stage, the embedding algorithm is implementedbased on the watermark bits by modifying coefficients, thatis, the LH or HL coefficient in every block will be altered.However, these binary bits need to be embedded into thefiltered blocks instead of random blocks as an effort to improvethe performance. The main idea of the proposed scheme is thatblocks with small differences are used to embed 0-bits and viceversa. Thus, they need to be sorted in the increasing of thedifferent values. It is important to note that the alteration onlymade for LH or HL coefficient. The algorithm for embeddingis represented as follow: ã  With watermark bit 0:   LH  i  =  LH  i  + ( y 1  − d 0 ,i ) ;( LH  i  ≥  HL i ) HL i  =  HL i  + ( y 1  − d 0 ,i ) ;( LH  i  < HL i )  (2) ã  With watermark bit 1:If   ( d 1 ,i  < y 2 )  with  ∆ =  y 2  − d 1 ,i   LH  i  =  LH  i  + ∆ HL i  =  HL i  −  ∆ ;( LH  i  ≥  HL i )   LH  i  =  LH  i  −  ∆ HL i  =  HL i  + ∆ ;( LH  i  < HL i ) (3)If   ( d 1 ,i  ≥  y 2 )   LH  i  =  LH  i HL i  =  HL i (4)where  LH  i  and  HL i  are the coefficients of LH and HL sub-bands of   i th sorted block.  d 0 ,i  and  d 1 ,i  are the differencesof the  i th block for 0-bit and 1-bit, respectively.  y 1  and y 2  are values need to be determined, in this paper, calledthe quantization threshold, which specifies the strength of watermarking. In above algorithm, it can be seen that thedifferences after modification will be  y 1  and  y 2  for 0-bit blocksand 1-bit blocks, respectively. The value of   y 1  is determinedfrom the differences of blocks which will be used for 0-bitsand computed as the following equation: y 1  =  Otsu ( d 0 ,i )  (5)where  Otsu  represents the Otsu method [12] which havebeen popularly applied in segmentation applications. With theprobability of data as the input, the Otsu method definesthe threshold based on minimizing the sum of within-classvariances correspondingto the sum of variations. It can be seenthat the total of errors after embedding for 0-bits depends onthe value of   y 1  in (2) and thus using the Otsu-based thresholdwill minimize the modification of coefficients. Unlike  y 1 , thevalues of   y 2  is determined in the different way as: y 2  =  d 1 ,i = λ × n 1  (6)where  n 1  is the number of 1-bits in the watermark. The valueof   y 2  is defined through the  λ  which represents the robustnessof our proposed watermarking.From (5) and (6), two thresholds are defined and presentedin the Fig. 2(b). In addition, it can be inferred that the differentvalues in blocks after embedding will be  y 1  and equal or larger                                              (a)                                             (b)                                             (c)Fig. 2. Different values of blocks of a sample image: (a) Probability of different value before embedding. (b) Determination of two thresholds. (c) Probabilityof different value after embedding          λ                        Fig. 3. Quality of output images in different values of   λ . than  y 2  as in the Fig. 2(c). Therefore, the embedded bits can bederived from this property in the extraction stage. Furthermore,the authors perform a simulation to evaluate the effect of   λ on the quality of the output images. As in the Fig. 3, somesample images have been investigated with different valuesof   λ . In particular, the PSNRs are reduced in increasing thevalues of   λ  for all of samples. We can see that the total of errors due to 1-bit embedding is enlarged when increasing thevalues of   y 2 . It is clear to know that the larger  λ  is, the more1-bit blocks have been altered. Thus, this is also the reason of the degradation of quality in the output images. The detail of a proposed embedding scheme is presented as in the Fig. 4.  B. The extracting algorithm In this stage, embedded bits need to be extracted based onthe comparisonof block differences and the adaptive threshold,denoted  y . The different value in each block of the embeddedimage can be only either  y 1  or larger than  y 2  as in the Fig.2(c). Therefore, an extraction is performed through the belowequation as: bit i  =   1 ( d i  ≥  y )0  otherwise  (7) Host Image 4-DWT Blocking Embedding Algorithm Unblocking4-IDWTEmbedded ImageWatermark ShuffleSortingRecoverKey (bit position) Fig. 4. The flowchart of an embedding process. where  d i  is the significant difference of   i th block of the em-bedded image and  y  is the extraction threshold. This thresholdneeds to be chosen in the range  ( y 1  < y < y 2 )  to ensure thatall of bits will be correctly recovered in both cases includingattacking or non-attacking. Thus, the problem in here is thathow to determine a threshold which can be used to extract toobtain the best performance under various critical attackingtypes. Moreover, this value needs to be adaptive for eachattacking type, that is, the thresholds are defined as differentvalues for different attacking types and different images. It isimportant to note that the accuracy rate is mostly dependedon this parameter, that is, it affects directly to the extractedwatermark quality. In this paper, the authors continuouslyapply the Otsu method to find out the threshold based onan effort of minimizing the number of wrong extracted bits.This work is performed by considering the probability densityfunction of differences of the embedded blocks under variouscases such as, non-attack, median filter  (7  ×  7) , rotation  5 0 ,and Gaussian noise (mean = 0, variance = 0.01) as shown inthe Fig. 5. However, the scattered distribution of blocks havingthe large differences can be affected to the computation of Otsu-based threshold. Thus, only blocks containing differenceslarger than the mean, denoted T  , will be modified to satisfy forcomputation. This step can be expressed through the followingequation: d i  =   d i  ( d i  ≤  T  ) T otherwise  (8)                                          (a)                                             (b)                                              (c)                                         (d)Fig. 5. Determination of threshold in different cases: (a) Non-attack. (b) Median filter  (7  ×  7) . (b) Rotation  5 0 . (d) Gaussian noise (mean = 0, variance =0.01) Embedded Image 4-DWT Blocking Extraction Algorithm Extracted Watermark Threshold    Key (bit position) Fig. 6. The flowchart of an extraction process. where  T  , the mean of differences, is computed as: T   = 1 N  N   i =1 d i  (9)where  N   is the total of blocks corresponding to the number of watermark bits. After determining threshold, bits are extractedby using (7). However, recovered bits need to be reshuffledby the key, containing the bit positions, which is createdin the embedding stage. Generally, the extraction scheme isperformed as the flowchart in the the Fig. 6.III. T HE  E XPERIMENTAL  R ESULT In this section, the proposed method are assessed in twostages. In the detail, the peak signal-to-noise ratio PSNR isused to estimated the image quality between the output andthe input. The formula to define PSNR is presented as follows: PSNR  = 10log 10   255 2 MSE    (10)where  MSE   is computed as below equation: MSE   = 1 M   × N  M   i =1 N   j =1 ( I   ( i,j )  − O ( i,j )) 2 (11)where  M   and  N   are the height and width of the image,respectively.  I   ( i,j )  and  O ( i,j )  are the grey value located atcoordinate  ( i,j )  of the srcinal image and embedded image. Inthe extraction process, the accuracy rate is expressed accordingthe normalized correlation NC value: NC   = 1 m × n m  i =1 n  j =1 w ( i,j )  × w ′ ( i,j )  (12)where  m × n  is the watermark of size.  w ( i,j )  and  w ′ ( i,j )  arethe values located at coordinate  ( i,j )  of srcinal watermark and extracted watermark. Value of   w ( i,j )  is set 1 if it isa watermark bit 1, otherwise, it is set (-1) and similar to w ′ ( i,j ) . So the value of   w ( i,j )  ×  w ′ ( i,j )  is either 1 or -1. If number of bits is extracted correctly, the NC value ispositive; otherwise, it is negative. Furthermore, the method isalso estimated under various aspects to consider the robustnessof a watermark based on the support from MATLAB.  A. Quality of embedded images In the below simulation, the authors use 10 common gray-scale images ( 512  ×  512  pixels, 8 bits/pixel), namely Lena,Baboon, Airplane, Pepper, Zelda, Goldhill, Couple, Elaine,Boat, and Truck from the USC-SIPI set; and the binary image( 32 × 32  pixels, 1 bits/pixel) as the watermark. The embeddedimages are shown in the Fig. 7 with the PSNR parameter of each sample. It is important to note that the wavelets of Haaris used for domain transformation and the strength factor  λ has been set at 0.4 as default. This value has been chosen tobalance an invisibility and a robustness of watermark.  B. Robustness of extracted watermark  To investigate the robustness of watermark under variousattacking types, the authors choose the  Lena  as the sample torepresent in the Fig. 8. These attacks have been divided intothree classes: geometric, non-geometric, and JPEG compres-sion class. Furthermore, results for all of tested images arealso provided in the Table. I.The geometric attacks include the cropping (remove quarterof the watermarked image and fill the missing portion with0-bits), scaling (resize to  (128  ×  128)  and then restored to (512  ×  512) ), rotation ( 5 0 and  25 0 ), Gaussian noise (mean=0,variance=0.01), Salt & Pepper noise (density=0.1), meanwhile,the histogram equalization, Gaussian filter  (3 × 3) , median, andaverage filter with different sizes of window are non-geometricattacks. Finally, the method is continuously estimated forJPEG compression with different levels through the QualityFactor, denoted QF. As in the Fig. 9, the watermarks areextracted from the attacked images. C. Compare to the others Finally, the authors employ the comparison test betweenthe proposed method and some recent approaches: Run [3]

Exp to Blood

Jul 23, 2017
Search
Similar documents
View more...
Tags
Related Search
We Need Your Support
Thank you for visiting our website and your interest in our free products and services. We are nonprofit website to share and download documents. To the running of this website, we need your help to support us.

Thanks to everyone for your continued support.

No, Thanks