Smithal and K.A. Navas'

Transcription

1 IEEE - ICSCN 2007, MIT Campus, Anna University, Chennai, India. Feb , pp Spatial Domain- High Capacity Data Hiding in ROI Images B. Smithal and K.A. Navas' Abstract: Digital watermarking, one of the data hiding techniques has become an emerging area of research due to the wide spread use of Internet and intranets. Though the watermark is used for authentication purpose, its methodology has been adapted for hiding data in many applications namely Electronic Patient Record (EPR) data hiding in medical images. Medical images are usually large sized images and stored without loss of redundancy. Recent researches have proven that appropriate level of JPEG (Joint Picture Expert Group) may be used on these image types without loss of diagnostic content. This has provided an opportunity for more rapid image transmission. This work focuses on the estimation of the data hiding capacity of Region of Interest (ROI) medical images and optimizing the JPEG survival level that allow acceptable JPEG for conventional spatial domain watermarking techniques namely LSB technique, Additive technique and Spread spectrum technique. EPR (Electronic Patient Report). The obtained results are very promising for the researchers in ROI watermarking arena such as medical and military image data hiding. The techniques used for ROI image watermarking are explained in Section II. Various performance metrics used to evaluate the effectiveness of watermarking are discussed in Section III. Section IV provides the implementation of three watermarking techniques and results. Finally, concluding remarks are provided in Section V. II. ROI IMAGE WATERMARKING TECHNIQUES In this section, we explain a typical data hiding scheme [3]. Let I be the original image, IROI be the image in which Region of Interest (ROI) is kept undistorted, data D represents the embedded information bit that is to be embedded. It is formulated as, (1) I =f(i,i, D) I. INTRODUCTION Data hiding techniques are emerging as an important area of research for various applications. One signal can be hidden in another signal imperceptibly [1, 2]. The signal can be one dimensional or multidimensional. For example a speech signal can be hidden in a speech, movie or image; an image can be hidden in an image, movie or speech. In this work, hiding text data in images is dealt with. More specifically, ASCII characters in ROI (Region of Interest) images. This work discusses the hiding of ASCII characters or text data as watermark in ROI images. At the receiver end, the watermarked image is separated into the original image and the ASCII characters. Original image and ASCII characters are given equal importance and taken care of. This work provides the estimation of capacity (Pay load) for various time domain watermarking techniques in ROI medical images. The pay load of LSB, additive and spread spectrum watermarking techniques for a specified ROI, quality factor and ratio are investigated. The pay load of the watermark is an important aspect in the hiding scenario of dat a (D) wateriiirlied itage (a) wateiiuwlred uhnage Recvery dta (D) Phase (b) Fig. 1. (A) Data Embedding (B) Data Recovery The visual quality of the modified image I should be close In data embedding method with the watermarked image I may be subjected to lossy such as JPEG, to that of original image IROI I =C() (2) where C(.) denotes the /de I is the watermarked operation, image after. Dept. of Electronics and Communication, College of Engineering, Trivandrum, Kerala, India. smitha d4yahoo.co.in, kanavas dgrediffmail.com /07/$ IEEE cover nge {I 528

2 IEEE-ICSCN, Feb Here is performed on the watermarked image to reduce the cost of storage and increase the speed of transmission. There are various spatial domain watermarking methods used for adding watermark and removing it from ROI images [10]. A. Least significant bit (LSB) watermarking technique In this method the least significant bit of each pixel element of the cover image is replaced by watermark information bit. Here, each character is represented by seven bits. At encoding stage, the given watermark which is a set of ASCII characters is converted into its binary equivalent; each binary bit is placed on the LSB of cover image. At decoding stage each binary bit is replaced and converted back to its ASCII equivalent to get the actual watermark. B. Additive watermarking technique In additive watermarking algorithm, the watermark and cover image are of same size. Consider a watermark D(m, n) embedded in cover image IROI (m, n) The basic embedding formula is, W(m, n) = IROI (m, n)(1+ ad(m, n)) (3) Where ac denote embedding strength and W(m, n) is the watermarked image. The retrieval can be done using the following expression D(m, n) = W(m, n) - IROI (in, n) (4) a x IROI (in, n) An alternate embedding formula is, W(m, n) = IROI (m, n) + acd(m, n) (5) The dewatermarking can be done using the following expression D(m, n)a where D(m, n) W(m, n) - IROI (m, n) is the original watermark recovered. C. Spread spectrum watermarking technique Consider watermark image D(m, n) as the watermark signal and PN sequence a(m, n) as the spreading signal. The desired modulation is achieved by applying both the watermark image and the PN sequence to a product modulator. The resultant signal I(m, n) is a pseudorandom (6) noise pattern that is added to the cover image IROI (m, n) to produce the resultant watermarked image W(m, n). The embedding formula is, W(m, n) = D(m, n)x a(m, n)+ IROI (m, n) (7) At the message recovery stage, watermarked image W(m, n) is correlated with the PN sequence which is an exact replica of that used for embedding the data. The unwanted noise signal can be filtered out during the process of correlation by setting the threshold as the mean value of correlation a. The watermark is said to be present if the correlation value exceeds the mean correlation. III. PERFORMANCE MEASURES The performance metrics used to evaluate the effectiveness of the ROI watermarking methods are discussed below. A. Data hiding capacity The data hiding capacity of a cover image is calculated as the maximum amount of information that can be embedded and recovered with low error probability. It is expressed in terms of number of message bits that can be embedded imperceptibly into each pixel of the specific cover image (bits per pixel). A simple approach to evaluate data hiding capacity is to model watermark channel as an Additive White Gaussian Noise (AWGN) channel [8]. According to AWGN model, watermark is added to a given set of features such as pixel intensities, transformed domain coefficients etc. from host image. The capacity of the channel is determined as, C =-og2 1 +'J bits/pixel (bpp) (8) where o 2 is the variance of watermark which denotes average energy per pixel allowed for the message. o 2 is the equivalent gaussian variance of the image noise, the image noise I are assumed to be uniformly distributed random variables taking values between 0 and 255. If we consider the channel noise due to two sources of noise [5], o-2 is the equivalent gaussian variance of the image noise and o2 is the variance of the processing noise. The p capacity of the data hiding channel can be expressed as, 529

3 Spatial Domain- High Capacity Data Hiding in ROI Images 1 K~~~~0 Ch7 1o2y± ± (9) h=2 log 2 (1 +, we calculate data hiding capacity assuming the distribution of the equivalent processing noise as gaussian distribution. The watermark capacity values for additive and spread spectrum methods are calculated using equation (8). In the calculation of watermark capacity, u2 is evaluated as the variance of watermark and o2 is evaluated as the variance of the cover image. B. Image imperceptibility The imperceptibility of a watermarking system refers to the perceptual similarity between the original and watermarked images. It is useful to measure the extent of distortion that the watermarking introduces to the work for the quality appraisal. The most conventional measures of perceptual image quality are Mean Squared Error (MSE) and Peak Signal to Noise Ratio (PSNR) techniques. Let original image be IROI (m, n) and the watermarked image W(m, n), an error function is defined as e(m, n) = IROI (m, n)- W(m, n) (10) The e(m, n) shows how close the watermarked image is to the original image. If e(m, n) equals to zero, that means no distortion introduced by embedding watermarks. Where larger e(m, n) exemplify more perceptible distortion. One of the simplest distortion measures is the Mean Square Error ( MSE or Erms ) function. I A17V,n M N E rims MSE = MNE E e(m, n) M=l n=1 (1 1) Where the image size is M by N, MSE is the mean of the squared error values across the entire image. The problem with MSE is that it depends strongly on the image intensity scaling. PSNR rectifies this problem by scaling the MSE according to the image range. Lm2 PSNR(dB) =10l1ogl0 MSE (12) where Lmax is the maximum value of luminance level (0-255). An alternative approach based on a model which tries to predict human observer's responses is Weighted Peak Signal to Noise Ratio (WPSNR). This is a different quality measurement suggested in [6]. The WPSNR uses an additional parameter called the Noise Visibility Function (NVF) which is a texture masking function. NVF uses a gaussian model to estimate how much texture exists in any area of an image. The WPSNR uses the value of NVF as a weighting factor. WPSNR(dB) 10I log10 1 (MExV Lmax MSE x NVF) (13) For flat regions, the NVF is close to 1 and for edge or textured regions NVF is more close to 0. The form of NVF is given as, NVF(m, n) 1 +O-J(MIn) IO(m,1 (14) where,2roj (inn) denotes the local variance of the image in a window centered on the pixel with coordinates (m, n) and 0 is a tuning parameter corresponding to the particular image. The image depend tuning parameter is given as, 0 : 2 D (7max IROJ (15) where 2maxIRo is the maximum local variance for a given image and D is an experimental value, range from 50 to 100. C. Compression effectiveness Compression effectiveness can be defined as the survival of the watermark given that an appropriate amount of JPEG is used. uncompressed image size Compression Ratio(CR)=cmrse iaesz compressed image size IV. EXPERIMENTAL RESULTS Cover images used are 8-bit gray scale images of bmp format with size 128X128. Three images were selected from each of the following modalities, Computed Tomography (CT), Magnetic Resonance Angiography (MA), Magnetic Resonance Imaging (MRI) and X-ray. ASCII characters or text images were used as watermarks. 4 percentage area of the cover image has been marked as ROI. In order to ensure imperceptibility of the watermark, WPSNR values are set to be >40dB. The data hiding capacity for additive and spread spectrum methods are calculated using equation (8) for two methods without and 530

4 IEEE-ICSCN, Feb with. A detailed discussion of the simulation results is given below. A. Least significant bit (LSB) technique A set of ASCII characters was used as the watermark. In the embedding stage, each ASCII character was encoded into seven bits. This binary information was embedded into LSB 1 plane of the cover image. The maximum number of bits that can be embedded (MNEB) into the LSB 1 plane was found to be bits. When binary information was embedded into both LSB1 and LSB2 planes, MNEB increased to bits. Similarly when embedded into LSB 1-3 planes, MNEB was found to increase to bits. Embedding data into the higher order bit planes resulted in perceptibility of the watermark. At the watermark recovery stage, the binary information in the LSB plane of the cover image was converted back to ASCII characters. The total number of characters embedded was 6744, which is equivalent to bits. It was observed for one of the CT image that, from LSB3 plane onwards, watermark insertion causes significant reduction in WPSNR values below the imperceptibility limit of 40 db. Therefore the MNEB obtained for this image was bits in LSB 1-2 planes. Table 1 shows the average capacity values attained for various modalities. Table 1: Average Values of MNEB for Each Modality (Without Compression) Modalithis LSB plane Capacity (bits) CT Eages MA images MRI imags X-raymimages In the method with, the watermarked image was subjected to JPEG. Although the resultant watermark was imperceptible, the recovered watermark was found to have errors. Therefore, techniques were not performed in LSB watermarking method. B. Additive watermarking technique In additive technique, the size of the cover image and watermark were taken the same. The watermark used was a text image of size 128X128. At the embedding stage, the watermark values were added to the cover image with a suitable value of embedding strength. The watermark was found to be imperceptible. The capacity was calculated using equation (8). The trade off between WPSNR and capacity is illustrated in Fig.2. The higher values of WPSNR resulted in reduced capacity. The average capacity values obtained for different modalities are summarized in Table2. Comparatively higher values of capacity were obtained for MRI images. P. nj WPSNR (db) CT lmri -e- X-ray lma Fig. 2. Reduction in Capacity with Increase in WPSNR (For the Method without Compression) Table 2: Average Values of Capacity Evaluated for the Method Without Compression Mo&U&s WPSNR (d) Capacity (bpp) CT images MA images MR images X-raymiages The watermarked image was subjected to JPEG for an image quality of 75. It was observed that relatively high ratio and capacity were obtained for MRI image. This may be due to large intensity variations in it. Watermark capacity of each modality was calculated by averaging the individual capacity scores of the images in that modality. The data hiding capacity of cover image was calculated in accordance with equation (8), taking watermark as the message signal and compressed cover image as noise signal. 1 I M.ii 1 A Oa4 rd 2 -L 4~~~~~~~~~~~~~~~~~~~~~~.5t3m compres non ratio 2Rg -t-ict Fig. 3. Variation in Capacity with Compression Ratio -Xr' -- MA Fig.3 gives a graph between ratio and average capacity. The watermark capacity was expressed in bits per pixel. i.e., the effective number of watermark bits i 531

5 Spatial Domain- High Capacity Data Hiding in ROI Images embedded on each pixel of the cover image. When the watermarked image is compressed; the redundancies in the cover images get reduced, resulting in increased number of watermark bits on each cover image pixel. It could be seen from Fig.3 that capacity increases with ratio. Table 4: Average Values for Capacity and WPSNR for Spread Spectrum Technique without Compression Table 3: Average Values of Capacity for Different Modalities, Evaluated for the Method with Compression. It can be seen that the Maximum Compression Ratio was Achieved for CT Images; Whereas Capacity and WPSNR Values are Maximum for MRI Images Modaties WPSNR (d) Capacity (bpp) CTimages MA images MRIimages X-ray imvs o0091 Table 5: Average Values of Capacity and WPSNR for Spread Spectrum Technique with Compression iodalitis Compression ratio WPSNR(dB) Capacity Opp) CT images MA mages CT images MEl images hia images Mil images X-ray images X-rayimages Modalities C. Spread spectrum watermarking technique In spread spectrum watermarking technique, the cover image was divided into blocks of size 8X8. A PN sequence was generated in accordance with a key. The objective of this technique was to hide each message bit into each block of the cover image. This was accomplished by the following steps. 1. If the message bit is zero, the PN sequence is added to the corresponding block of cover image. 2. On the other hand, if it is one, the corresponding block of cover image is left unchanged. At the watermark recovery stage, the same PN sequence was generated using the key. Each one of the 8x8 block of watermarked image was correlated with the PN sequence. The message bit was decoded as 0 for the blocks having high values of correlation with PN sequence. Otherwise, it was decoded as 1. The watermark used in this technique was a text image of size 12X9. It was embedded into cover images of different modalities. Then the WPSNR and capacity values were evaluated for different modalities. The average WPSNR and capacity values calculated for without method are tabulated in Table4. It is obvious that the WPSNR values obtained are much lower than the minimum required value of 40dB. The watermarked image was compressed for an image quality of 75. It was observed that there is a noticeable difference between cover image and the compressed watermarked image. The average values are tabulated in Table5. These values were found to be smaller compared to the results obtained for additive and LSB techniques. Compression WS-R (db) Capaciy (bpp) io D. Comparison of data hiding capacity The data hiding capacity of three different spatial domain watermarking techniques such as least significant bit, additive and spread spectrum were compared and the results are summarized in Table Table 6: Comparison of Data Hiding Capacities of LSB, Additive and Spread Spectrum Watermarking Techniques Capacity in number of bits Modalities LSB Additive method r-sa spectrum method Spread Without With WV ithout -- CT images MA images MRI images X-ray images The LSB technique was found to have a maximum data hiding capacity of bits, when data was embedded into LSB 1-3 planes. The CT images were found to have a reduced average capacity value because data can be embedded imperceptibly only in LSB 1 and LSB2 planes for one of these CT images. Using additive method with, a maximum capacity of bits was obtained for MRI images. It was

6 IEEE-ICSCN, Feb observed that the additive method attained maximum capacity when data was embedded in a single LSB plane. Lowest capacity values were obtained for spread spectrum method. The maximum data hiding capacity attained for this method was 234 bits in MRI images using technique. Among the cover images of different modalities, maximum data hiding capacity was obtained for MRI images in all the three watermarking techniques. This may be due to the large intensity variations in the image. V. CONCLUSION Three spatial domain watermarking techniques namely LSB, additive and spread spectrum were implemented. Data hiding capacities for these techniques were evaluated and compared. The maximum data hiding capacity was obtained for LSB technique. Attempts were made to measure the capacity of cover image if they are watermarked and compressed. LSB method was found to be not tolerant to. Watermark could not be recovered from the watermarked image using LSB technique when it is subjected to. Watermarked image obtained using additive method could be compressed more than that obtained using spread spectrum technique for the given pay load. REFERENCES [1] Stefan Katzenbeisser and Fabien A.P.Petitcolas, "Information Hiding Techniques for Stegnography and Digital Watermarking," Artech house, Computer security series, pp , , [2] Neil F.Johnson, Zoran Duric and Sushil Jajodia, "Information Hiding Steganography and Watermarking - Attacks and Counter Measures," Kluwer academic publisher, pp , [3] Kiranmayi Penumarthi, "Augmented Watermarking," Thesis, Master of Science in Electrical Engineering, Louisiana State University and Agricultural and Mechanical College, pp , December [4] Dominic Osborne, " Embedded Watermarking for Image Verification in Telemedicine," Thesis, Doctor of Philosophy in Electrical and Electronic Engineering, University of Adelaide, pp , [5] Mahalingam Ramkumar and Ali N.Akansu, "Theoretical Capacity Measures for Data Hiding in Compressed Images," SPIE Symposium on Voice, Video and Data Communication, Boston MA, vol. 3528, pp , November [6] Qi and Pei, "Impact Analysis of Digital Watermarking on Perceptual Quality using HVS Models, "Graudate in Electrical and Computering Engineering, University of Wisconsin, Madsion. [7] Ruizhen Liu and Tieniu Tan, "Theoretical Framework for Watermark Capacity and Energy Estimation," Media Sce Technologies GmbH, Essen, Germany. [8] T.M.Cover and J.A.Thomas, "Elements of Information Theory," Second Edition, John Wiley and Sons Inc, pp , [9] C.Y. Lin and S.-F. Chang, "Zero-error Information Hiding Capacity of Digital Images," submitted to IEEE Intl. Conf. on Image Processing, Oct [10] Chris Shoemaker, "Hidden Bits: A Survey of Techniques for Digital Watermarking," shoemakc/watermarking/ watermarking.html 533