Effect of Watermarking in Vector Quantization based Image compression

Transcription

1 Effect of ing in Vector Quantization based Image compression Soumyo Bose Madhulika Suvojit Acharjee Dept. of ETCE, Jadavpur University Kolkata, West Bengal, India Shatadru Roy Chowdhury Sayan Chakraborty Nilanjan Dey Dept. of IT, JIS College of Engineering Abstract In the modern era of health care and medical diagnostics, medical image contents are needed to be transported through the internet, in a safe and secured way, within time. Image compression is a process of reducing the size of an image without affecting the quality. Image compression reduces the size and helps in quick transportation. Digital watermarking is a process of embedding some secret data in an image for copyright protection, authentication, and channel reliability detection, done without affecting its look. In this work, vector quantization based compression technique is used. It follows LBG clustering algorithm. For watermarking, DCT-DWT based embedding and extraction technique is adopted. At first, retina image was compressed and BPP, SNR, MSE, PSNR values were compared, before and after compression. Later, a binary EPR logo image was embedded into compressed and uncompressed images. Correlation values of original watermark and extracted watermark images were computed and compared for further study. Keywords DCT, DWT, Vector Quantization, Codebook, BPP, MSE, SNR, Correlation. I. INTRODUCTION We know that, from ancient age, images are used for communication. Since past few decades, advancement of multimedia applications and vast use of the web has led to the increase in need for processing of images. In order to reduce internet load and to conform the channel capacity, image compression is required. More precisely, image compression is minimizing the size of a graphics file in bytes without degrading the quality of the image to an unacceptable level. The reduction in file size allows more images to be stored in a given amount of disk or memory space. It also reduces the time required for images to be sent over the internet or downloaded from web pages. A text file or program can be compressed without the introduction of errors, but only up to a certain extent. This is known as lossless compression, which is of primary importance to the medical imaging [1]. Beyond this point, errors are introduced. In text and program files, it is crucial that compression be lossless because a single error can seriously damage a text file, or stop a program execution. In image compression, a small loss in quality is usually not noticeable. There is no "critical point" up to which compression works perfectly, but beyond which it becomes impossible. If there is some tolerance for loss, then the compression factor can be greater. For this reason, graphical images can be compressed more than text files or programs. Vector quantization (VQ) is a classical quantization technique. It works by dividing a large set of vectors into groups having approximately the same number of vectors closest to them. Each group is represented by its centroid point as in k-means and other clustering algorithms. It is a lossy data compression method on the principal of block coding. VQ is a fixed-to-fixed length algorithm. It has been applied to various applications like pattern recognition [2], density estimation, data compression [3], face detection, image segmentation [4], speech recognition [5],tumor detection [6] etc. VQ can be defined as a mapping function that maps k-dimensional vector space to a finite set CB = {C1, C2, C3,.., CN}. The set CB is called codebook consisting of N number of code vectors and each code vector Ci = {ci1, ci2, ci3,,cik} is of dimension k. The key to VQ is the good codebook [7]. The method most commonly used to generate codebook is the Linde-Buzo-Gray (LBG) algorithm, which is also known as Generalized Lloyd Algorithm (GLA), introduced by Yoseph Linde, Andres Buzo and Robert M. Gray in 1980 [8]. It has some similarity with k- means clustering [9] algorithm. In the area of medical information exchange, transferring patient s valuable data via internet or any other medium from one organization to the other, copyright protection[10] from potential threats and authentication[11] of data is highly necessary. Digital data, such as bio-medical images, audio, video files are needed to be secured by embedding digital watermarks into them. ing [12,13,14] not only 547

2 helps in data security, but also helps to check the channel reliability by tampering detection [15], and helps to add extra information to the multimedia file. ing [16, 17] is mainly of three types - robust watermark [18] that resists all sort of deformations and compressions; semi-fragile watermark that resists benign transformations, but not malignant; fragile watermark that cannot resist slightest of deformation and it is mainly used for tamper detection. II. METHODOLOGY A. LBG Clustering Algorithm: One of the most popular clustering [19] algorithms is the LBG algorithm. The LBG clustering algorithm is the simplest and most commonly used in vector quantization. It s clustering is done on the basis of codebook size. Initially, it determines the size of codebook, say N. Then it selects N random points as initial centroid and cluster input vectors into N region. For, each and every vector it measures Euclidean distance and on that basis it updates its codebook. Step 1. Size of the codebook is set as N, where N= 2 k (k=1,2,3,..). Step 2. N codewords are selected at random from the set of input vector. Step 3. The vector around each code word clustered using Euclidean distance is measured. Step 4. The minimal distortion is found. Step 5. New codebook is computed and the previous one is updated. Step 6. Steps 3 and 4 is repeated until final codebook is generated. Distortion Measure: ( l 1) ( l) D D ( l 1) D (1) where, D is distortion, l is number of iteration and threshold value is represented by ɛ For this algorithm, initial requirement is the twodimensional vector space. Then it requires setting cluster number, say N. Now the vector space is partitioned into N clusters. It repeatedly finds the centroid of each set in the partition, and then re-partitions the input vector according to which these centroids are closest as shown in Fig 1. Fig 1. LBG clustering B. DCT-DWT ing DCT (discrete cosine transformation) [20] is a popular function in signal processing where a signal is transformed from spatial domain to frequency domain. DCT divides an image into m x n blocks where each block consists of 8 x 8 components. If an input image be x, then its DCT components are computed as per the following equation M 1N 1 (2 m 1) u (2n 1) v y(u, v) u v x(m, n) cos cos M N u 0 v 0 2M 2M Where, u = v = 1 u=0 2 1 u=1, 2,, N-1 1 v=0 2 1 v=1, 2, N-1..(1) DCT transformation applied to each of the aforesaid blocks of an image converts the input image into three frequency sublayers, low-frequency, mid-frequency and high-frequency subbands. Human vision is more sensitive to low-frequency subband. High-frequency sub-band is removed by compression procedure. That leaves the mid-frequency band where watermarking is applied, which means the original image can neither be deformed, nor the watermark be removed by compression. DWT (discrete wavelet transformation) [21] is a singlelevel discrete 2-D wavelet transformation method. DWT divides the input image into four low-resolution components, the approximation coefficient matrix (LL), the horizontal detail matrix (LH), the vertical detail matrix (HL) and the diagonal detail matrix (HH). 548

3 LL 1 LL 2 HL 2 LH 2 HH 2 LH 1 HH 1 Fig 2. 2-level DWT method The process can be repeated to compute multiple scale decomposition. Embedding [22] in any of the three detail band allows increasing the robustness of watermark [23, 24] without affecting the image quality. III. PROPOSED METHOD The steps of the proposed method are as follows: Step 1. A medical gray retina image is taken for compression using vector quantization method. Step 2. LBG algorithm is applied by setting codebook-length (32, 64 and 128). Step 3. Compression ratio, BPP, SNR, MSE, PSNR for original and compressed images are calculated. Step 4. Gray retina image and compressed gray retina image are chosen as cover. Step 5. Binary watermark is embedded into the cover frames. Step 6. Binary watermark is extracted from compressed images and non-compressed images. Step 7. Error images between the extracted watermark images before and after compression are obtained. IV. EXPLANATION OF THE PROPOSED METHOD The proposed method first takes a medical gray retina image. Then vector quantization is applied for compression of image. Here LBG algorithm was used for compression. In LBG algorithm, first codebook size is set as 32, 64 and 128. Then, 32, 64 and 128 random codewords are chosen from the input vector and the image is clustered in 32, 64 and 128 centroid. Then in next iteration, the distance from the centroids is calculated. Vectors are assigned to that centroid which is closest and the codebook is updated. This process is repeated until we get the final codebook with minimum distortion. Then compression ratio, original and compressed BPP, SNR, MSE, PSNR is calculated for compressed image. In our previous research [25], we have discussed the technique of watermark embedding and extraction using DWT-DCT. The DWT is applied up to level 2 in compressed image and gray retina image to decompose the host image into four sub bands: LL2, HL2, LH2, and HH2. The purpose of applying DWT up to level-2 was to get four smaller sub-bands & HL 2 or HH 2 was chosen for DCT application. HL 2 (or HH 2 ) sub-band is divided into 4 x 4 sub-blocks. DCT was applied in each HL2 or HH2 sub-bands. Thus, the grayscale watermark image is transformed into a vector of zeroes and ones, means binary watermark image. Two uncorrelated pseudorandom sequences were generated. One sequence is used to denote the watermark bit 0 & the other sequence is used to denote the watermark bit 1. Then, binary watermark is hidden into the DCT sub bands using pseudo-random sequences. Two pseudorandom sequences are fixed with a gain factor. Thereafter, inverse DCT is applied to generate watermark image and again inverse DWT is applied to produce the watermark host image. The process is shown in the figure below. Original image 2-Scale DWT DCT Embedding algorithm IDCT 2-level IDWT ed Image formation Pseudo numeric sequence Key Fig 3. Embedding Procedure Finally, watermark image was extracted from the compressed as well as uncompressed images. The extraction process was quite similar to that of embedding process. 2- scale DWT [25] and then DCT were applied to the watermarked images. Two sequences of pseudo random numbers were generated. Finally, with the help of binary watermark size in the recipient end the watermark is extracted from DCT sub-band followed by DWT sub-bands. The process is shown in the figure below. ed Image 2-Scale DWT DCT Extraction algorithm Extracted formation Pseudo numeric sequence Fig. 4. extraction process For this study, 512 X 512 gray retina images were used as cover image and 64 X 64 binary images were used as watermark image. For compression codebook size is selected as 32, 64 and 128 on the basis of image size. For large images codebook size may be 256 or 512. So, codebook size must be fixed on the basis of image size for better performance. 549

4 Mean-square Error Measure: MSE= Peak signal-to-noise ratio Measure: 1 ˆ n 2 ( Yi Yi) (2) n i 1 codebook _ 2 PSNR=10* log ((2 length 1) / MSE) (3) 10 Bit per Pixel Measure: BPP = Total number of bits in final file/ number of pixels in final file Compression Ratio Measure: CR = total number of bits in final file/ number of bits in original file Where, ˆ Y is a vector of n predictions and Y is the vector of the true values. V. RESULTS AND DISCUSSION The complete study includes image compression, watermark embedding and extraction of watermark image from compressed and uncompressed images. For image compression, watermark embedding [26, 27] and extraction, MATLAB R2013a software was used. The proposed method showed remarkable results as an effect of compression and watermarking [28, 29, 30]. To show our proposed method s efficiency, we have tested this method by using different codebook length. First we watermarked it without using compression and then by using compression. Different results are produced for both and for different codebook length. The results are shown in Table I and Table II. Codebook Size Table I. STATISTICAL MEASUREMENT OF VQ COMPRESSION CR Original BPP Compressed BPP SNR MSE PSNR CR BPP MSE A Codebook 128 Codebook 64 Codebook Correlation Uncompressed CB=128 CB=64 CB= SNR Fig. 5 (A) Comparative study of CR, BPP and MSE for CB size 128,64 and 32, (B) Comparative study of SNR,PSNR for CB size 128,64 and 32. Table II. CORRELATION OF EXTRACTED WATERMARK IMAGES WITH THAT OF ORIGINAL WATERMARK Image PSNR Codebook 128 Codebook 64 Codebook 32 Correlation Extracted watermark, uncompressed Compressed extracted watermark, CB= Compressed extracted watermark, CB= Compressed extracted watermark, CB= B Fig.6 Comparative study of Correlation for uncompressed CB size 128, 64 and 32, Table I. shows statistical measurement of VQ compression and Fig. 5 shows comparative study of CR, BPP, SNRMSE and PSNR for CB size 128, 64, 32. From Table I. it can be concluded that by increasing codebook size we can increase compression ratio as well as BPP, i.e. we can get more reduced image. It also shows from Table I, that, by increasing codebook size we can reduce mean-square-error and can increases PSNR and SNR, i.e. we can get less noisy images; but we have to fix the codebook size depending upon the image size, so that we can get more compressed noise free images. From Table II it can be concluded that correlation decreases by using compression technique. But, we can fix it by increasing codebook size. As Fig. 6 shows, by fixing codebook size to 64 and128, we can get very close correlation as compared to uncompressed image. So we have to fix codebook size, as effect of compression on correlation to become less. 550

5 A D G A B C D Fig. 7.(A) Original Image, (B) ed Image, (C) Original watermark, (D) Extracted watermark B E H Fig.8. (A) Compressed image for CB size=128, Fig. 8 (B) Compressed watermarked image for CB size=128, Fig. 8 (C) Extracted watermark for CB size =128, Fig. 8 (D) Compressed image for CB size =64, Fig. 8 (E) Compressed watermarked image for CB size =64, Fig. 8 (F) Extracted watermark for CB size=64, Fig. 8 (G) Compressed image for CB size=32, Fig.8 (H) Compressed watermarked image for CB size=32, Fig. 8 (I) Extracted watermark for CB size=32. Fig. 7 shows the effect of watermarking on uncompressed gray retina image and Fig. 8 shows the effect of compression using CB size 128, 64, 32 on watermark image. Extracted watermark images are shown in Fig. 8 (G), Fig. 8 (F) and Fig. 8 (I) for CB size 128, 64 and 32 respectively. VI. CONCLUSION The proposed study was about the effect of watermarking in image compression on medical image contents. In the present study, medical gray retina image was compressed using vector quantization method. Statistical measures of image compression were taken for comparative study. Then watermark image was embedded and extracted from both compressed and uncompressed images. From Table II, it is clear that when codebook length increases, correlation between watermark and extracted watermark increases. In this project, vector quantization method was used for image compression. There are other available watermarking and image compression methods which can be used in further studies. C F I REFERENCES [1] A.C. Flint, Determining optimal medical image compression: psychometric and image distortion analysis, BMC Medical Imaging [2] F. Ma, M. Bajger, J. P. Slavotinek, and M. J. Bottema, Two graph theory based methods for identifying the pectoral muscle in mammograms, PatternRecognition., vol. 40, no. 9, pp , Sep [3] K.Zeger and A. Bist, Universal adaptive vector quantization using codebook quantization with application to image compression, vol.3, ISBN: , [4] Q.Kanafani, A. Beghadadi and C. Fookes, Segmentation-based image compression using BTC-VQ technique, Signal Processing and Its Applications, vol.1, pp: ,2003. [5] J.Makhoul, S. Roucos and H. Gish, Vector quantization in speech coding,vol:73, ISBN: , [6] H.B.Kekre and T.K. Sarode, Detection and demarcation of tumor using vector quantization in MRI images, International Journal of Engineering Science and Technology, vol.1, pp-59-66, [7] H.B. Kekre and T.K. Sarode, 2-level vector quantization method for codebook design using Kekre s median codebook generation algorithm, Advances in Computational Sciences and Technology ISSN vol.2, pp , [8] Y.Linde,A.Buzo and R.Gray, An algorithm for vector quantizer design,vol:28, ISSN: ,1980. [9] S.N.Sulaiman and N.A.M. Isa, Adaptive fuzzy-k-means clustering algorithm for image segmentation, Consumer Electronics, IEEE Transactions, pp: , [10] R.M. Prasad and S. Koliwad, A robust wavelet-based watermarking scheme for copyright protection of digital images,computing Communication and Networking Technologies (ICCCNT), pp: 1 9, [11] N. Dey, B. Nandi, P. Das, A. Das and S.S. Chaudhuri, Retention of electrocardiogram features insignificantly devalorized as an Effect Of ing For a multi-modal biometric authentication system, Advances in Biometrics for Secure Human Authentication and Recognition, Taylor and Francis, pp. 450, July 2013, (in Press). [12] C. T. Hsu and J. L. Wu, Hidden digital watermarks in images, IEEE Trans. On Image Processing, vol. 8, no. 1, pp , Jan., [13] M. Barni, C.I. Podilchuk, F. Bartolini and E.J. Delp, Embedding: hiding a signal within a cover image. Communications Magazine, IEEE volume: 39, pp: , [14] J. Nin and S. Ricciardi, Digital watermarking techniques and security issues 27th International Conference on Advanced Information Networking and Applications Workshops (WAINA), pp , [15] N. Dey, M. Dey, D. Biswas, P. Das, A. Das and S.S. Chaudhuri, Tamper detection of electrocardiographic signal using watermarked bio-hash code in wireless cardiology, Special Issue of the International Journal of Signal and Imaging Systems Engineering,Inderscience, 2013 (in Press). [16] A.K. Pal, P. Das and N. Dey, Odd-Even embedding scheme based modified reversible watermarking technique using Blueprint - arxiv preprint arxiv: , [17] N.Dey, S.Chakraborty and S.Samanta, Optimization of watermarking in biomedical signal, [Lambert Publication],Heinrich-Böcking-Straße 6, Saarbrücken, Germany],2013. [18] S. P. Singh, P. Rawat and S. Agarwal, A robust watermarking approach using DCT-DWT, International Journal of Emerging Technology and Advanced Engineering, vol. 2, Issue 8, August [19] H.B.Kekre and T.K. Sarode, New clustering algorithm for vector quantizationusing rotation of error vector, International Journal of Engineering Science and Technology,vol.7, [20] N.Dey, M.Pal and A.Das, A session based watermarking technique within the NROI of retinal fundus images for authentication using DWT,spread spectrum and Harris corner detection, International Journal of Modern Engineering Research, vol.2, Issue 3,pp ,

6 [21] C. Wang, C. Zhang,P. Hao, A blind video watermark detection method based on 3d-DWT transform, 17th IEEE International Conference on Image Processing (ICIP), [22] S. Chakraborty, P. Maji, A.K. Pal, D. Biswas and N. Dey, Reversible Color Image ing using Trigonometric Functions, International Conference on Electronic Systems, Signal Processing and Computing Technologies, pp , [23] N. Dey, P. Maji, P. Das, A. Das and S.S. Chaudhuri, An Edge Based ing Technique of Medical Images without Devalorizing Diagnostic Parameters, International Conference on Advances in technology and Engineering, pp.1-5, 2013 [24] N. Dey, P.Das, A. Das and S.S. Chaudhuri, DWT-DCT-SVD Based Intravascular Ultrasound Video ing,second World Congress on Information and Communication Technologies (WICT 2012), pp , [25] N. Dey, P. Das, A. Das and S.S. Chaudhuri, DWT-DCT-SVD Based Blind ing Technique of Gray Scale Image in Electrooculogram Signal, International Conference on Intelligent Systems Design and Applications (ISDA-2012), pp , [26] N. Dey, S. Biswas, A. B. Roy, A. Das and S.S. Chaudhuri, Analysis Of Photoplethysmographic Signals Modified by Reversible ing Technique using Prediction-Error in Wireless Telecardiology, International Conference of Intelligent Infrastructure, 47th Annual National Convention of CSI,2012, McGraw-Hill Proceeding. [27] S. Chakraborty, S. Samanta, A. Mukherjee, N. Dey and S.S. Chaudhuri Particle Swarm Optimization Based Parameter Optimization Technique in Medical Information Hiding, 2013 IEEE International Conference on Computational Intelligence and Computing Research(ICCIC), pp. 1-6, [28] N. Dey, P.Das, D. Biswas, P. Maji, A. Das and S.S. Chaudhuri, Visible ing within the Region of Non-Interest of Medical Images Based on Fuzzy C-Means and Harris Corner Detection, The Fourth International Workshop Communications Security & Information Assurance (CSIA-2013)[Springer], pp , [29] N. Dey, S. Biswas, P. Das, A. Das, and S.S. Chaudhuri, Lifting Wavelet Transformation Based Blind ing Technique of Photoplethysmographic Signals in Wireless Telecardiology, Second World Congress on Information and Communication Technologies (WICT 2012), pp , [30] N. Dey, S. Biswas, P. Das, A. Das, and S.S. Chaudhuri, Feature Analysis for the Reversible ed Electrooculography Signal using Low Distortion Prediction-error Expansion, 2012 International Conference on Communications, Devices and Intelligent Systems (CODIS), pp ,