IEEE SIGNAL PROCESSING LETTERS, VOL 18, NO 2, FEBRUARY 2011 95 A New Histogram Modification Based Reversible Data Hiding Algorithm Considering the Human Visual System Seung-Won Jung, Le Thanh Ha, and Sung-Jea Ko Abstract—In this letter, we propose an improved histogram modification based reversible data hiding technique In the proposed algorithm, unlike the conventional reversible techniques, a data embedding level is adaptively adjusted for each pixel with a consideration of the human visual system (HVS) characteristics To this end, an edge and the just noticeable difference (JND) values are estimated for every pixel, and the estimated values are used to determine the embedding level This pixel level adjustment can effectively reduce the distortion caused by data embedding The experimental results and performance comparison with other reversible data hiding algorithms are presented to demonstrate the validity of the proposed algorithm Index Terms—Data hiding, human visual system, just noticeable difference, lossless watermarking I INTRODUCTION R EVERSIBLE data embedding, which is often referred to as lossless or invertible data embedding, is a technique that embeds data into an image in a reversible manner In many applications including art, medical, and military images, this reversibility is a very desirable characteristic, and thus considerable amount of research has been done over the last decade [1]–[6] In the conventional works, extensive efforts have been devoted to increase the embedding capacity without deteriorating the visual quality of the embedded image A key of reversible data embedding is to find an embedding area in an image by exploiting the redundancy in the image content Early reversible algorithm [1] uses lossless data compression to find an extra area that can contain to-be-embedded data In order to expand the extra space, the recent algorithms reduce the redundancy by performing pixel value prediction [2]–[5] and/or utilizing image histogram [5], [6] The state-of-the-art techniques [4], [5] exhibit high embedding capacity without severely degrading the visual quality of the embedded result Manuscript received October 06, 2010; revised November 16, 2010; accepted November 20, 2010 Date of publication December 03, 2010; date of current version December 20, 2010 This work was supported by Korea University Grant, Seoul Future Contents Convergence (SFCC) Cluster established by Seoul R&BD Program (10570), and Mid-career Researcher Program through National Research Foundation of Korea (NRF) grant funded by the Korea government (MEST) (2010-0000449) S.-W Jung and S.-J Ko are with the Department of Electrical Engineering, Korea University, Seoul, Korea E(e-mail: jungsw@dali.korea.ac.kr; sjko@korea.ac.kr) L T Ha is with the Department of Information Technology, University of Engineering and Technology, Vietnam National University, Hanoi, Vietnam (e-mail: ltha@dali.korea.ac.kr) Digital Object Identifier 10.1109/LSP.2010.2095498 However, since the subjective visual quality is not taken into account in the conventional methods, the quality of the resultant embedded images is often not satisfactory In this letter, we propose a histogram modification based reversible data embedding algorithm considering the human visual system (HVS) In the proposed algorithm, a local causal window is used to predict a pixel value and estimate an edge Then, by taking a concept of the just noticeable difference (JND) [7], [8], the pixels in the smooth and edge regions are differently treated to reduce the perceptual distortion Experimental results demonstrate that as compared to conventional algorithms, the proposed algorithm produces subjectively higher quality embedded images while providing a similar embedding capacity This letter is organized as follows In Section II, the proposed scheme is described The performance of the proposed algorithm is evaluated and compared with the conventional algorithms in Section III, and finally, we conclude the paper in Section IV II PROPOSED ALGORITHM The proposed algorithm is based on histogram modification The conventional histogram modification methods embed a message bit into the histogram of pixel values [6] or the histogram of the pixel differences [5] Since Tai et al.’s method [5] outperforms Ni et al.’s method [6], Tai et al.’s method is chosen as our basic framework Compared to Tai et al.’s work, our major contribution is to use a causal window to predict a pixel value, an edge, and the JND, and to exploit these predicted values when performing data embedding Let and denote the original and the embedded image, , the pixel value is respectively For each pixel coordinate predicted by (1) where represents a causal window surrounding and returns a cardinality of the set For instance, the shown in Fig contains 12 pixel causal window of size positions and the average of the pixel values at these positions is used as a predicted value Then we calculate the pixel difference between the original and predicted values by (2) where process 1070-9908/$26.00 © 2010 IEEE is the difference value used in the data embedding 96 IEEE SIGNAL PROCESSING LETTERS, VOL 18, NO 2, FEBRUARY 2011 to-be-embedded bits Thus, the overflow and underflow problem can happen when the embedded value exceeds a pixel value bound (0 to 255 in bit images) To solve this problem, the original image histogram is shrunk from both sides by , where is the embedding level To realize reversible data embedding, the overhead information describing this preprocessing is losslessly compressed and embedded together with pure payload data Detailed description for preprocessing can be found in [5] For the notational simplicity, from now on, let denote the preprocessed version of the original image and then (1)–(4) are applied to the preprocessed image Unlike Tai et al.’s method adopting the fixed embedding level , we adaptively adjust the embedding level for each pixel according to the local image , the characteristics For the non-edge pixel, i.e., if is defined by embedding level Fig Causal window for computing E (i; j ), Jnd(i; j ), and x ^(i; j ) (5) In other words, a maximum possible embedding level is chosen with a constraint that the pixel value change should be lower than the JND value This is because the distortion above the JND in the smooth region is perceptually disturbing On the other , is determined by hand, for the edge pixel of (6) Fig Visibility threshold against background luminance [7] In the proposed method, the perceptual characteristic of the HVS is exploited to alleviate the quality degradation caused by data embedding To this end, the edge is simply estimated for each pixel as follows: if if (3) where indicates whether the pixel is the edge or not, represents the variance of pixel values in , and is an edge threshold Since the HVS is known to perceive the difference above the JND, the JND value is estimated after edge detection as follows: (4) are two thresholds representing the luminance where and adaptation and the activity masking of the HVS characteristics, [7] In order to estimate , backrespectively, and ground luminance is first measured by taking the average value of the local neighborhood Then, a piecewise linear approximation in Fig is used with three parameters, , , and , described , , and for nonedge in [7] Specifically, pixels (i.e., when ), and , , and for edge pixels (i.e., when ) In addition, is defined as the maximum pixel difference value in the local neighin (3), background borhood Note that when computing luminance, and in (4), only the pixels in the causal window are used because only these pixels are available at the data extraction stage due to the raster scan order processing Actual data embedding is performed by increasing the differand finding the extra space that can contain ence value Namely, a minimum possible embedding level above the JND is used to embed a sufficient amount of data This is because it is difficult to find the extra space using the embedding level lower than the JND since the difference values in the edge region are high Besides, the increase of the JND in the edge region does not severely deteriorate the visual quality and sometimes an intentional increase of the JND in the edge region is employed in the image enhancement algorithm [9] After estimating the edge, the JND, and finally the embedding level, we can try to embed a message bit for each pixel If , the message bit is embedded by (7) Otherwise, if , data embedding is not performed but the difference value should be expanded to discriminate this pixel from the embedded pixels In this case, the output pixel value is obtained by (8) Since the pixel value is shifted by at preprocessing, underflow or overflow is prevented This data embedding process continues until all to-be-embedded bits are inserted and the resultant embedded image can deliver the embedded information Given only the embedded image and the embedding level , the original image is recovered and the embedded bits are obtained at the data extraction process Because the pixel value prediction, edge and JND computation, and embedding level estimation are performed by using the causal window, the same values can be derived at the extractor Here, the pixels at the JUNG et al.: NEW HISTOGRAM MODIFICATION 97 Fig Test images of 256 256 bits In the upper row and from left to right: Airplane, Baboon, Boat In the lower row from left to right: Candy, Lena, Peppers Fig SSIM versus watermark capacity for test images upper and left image boundaries are not modified to satisfy the reversibility , the At the data extractor, if message bit is extracted by if if is even is odd (9) , the pixel value is re- Then, when covered by if if (10) represents ceiling operation where Otherwise, if is compensated by , the shifted value if if (11) Since the overhead information bits describing the preprocessing are also extracted, the original image is finally recovered by shifting back the image histogram III EXPERIMENTAL RESULTS In order to evaluate the performance of the proposed algorithm, six commonly used grayscale images shown in Fig are used [10] First, the capacity versus distortion performance of the proposed algorithm is illustrated in Fig Here, the distortion is measured by the structural similarity (SSIM), which effectively assesses the perceptual visual quality of the image [11], and the capacity is represented by the average number of and the embedded bit per pixel (bpp) The edge threshold causal window size are empirically determined by 200 and 3, respectively For all test images, more bits can be embedded by increasing the embedding level at the expense of the quality degradation Because data embedding is dependent on the redundancy in the image content, images containing a large smooth area such as Fig Performance comparison among Tai’s, Hu’s, and proposed methods for the Lena image: (a) watermark capacity versus PSNR, (b) watermark capacity versus SSIM Candy can embed a large number of bits, whereas images with complicated textures such as Baboon can contain a relatively small number of bits 98 IEEE SIGNAL PROCESSING LETTERS, VOL 18, NO 2, FEBRUARY 2011 suitable to the conventional applications of the reversible data hiding, such as art, medical, and military imaging In addition, note that the proposed algorithm produces the embedded images exhibiting sharper image details compared to the original images Therefore, even though the image enhancement is not a concern in reversible data hiding, the embedded image can replace the original image in some applications, where the sharp image details are preferred In such applications, the proposed algorithm can be used to perform the image enhancement and reversible data hiding at the same time IV CONCLUSION Fig Magnified regions of the original (first column) and watermarked (second to third columns) images: (a) Lena, (b) Tai’s (PSNR: 28.98 dB, SSIM: 0.860, capacity: 0.829 bpp), (c) the proposed (PSNR: 31.04 dB, SSIM: 0.922, capacity: 0.802 bpp), (d) Peppers, (e) Tai’s (PSNR: 29.70 dB, SSIM: 0.861, capacity: 0.885 bpp), (f) the proposed (31.03 dB, SSIM: 0.915, capacity: 0.807 bpp), (g) Airplane, (h) Tai’s (PSNR: 34.20 dB, SSIM: 0.927, capacity: 0.776 bpp), (i) the proposed (PSNR: 34.02 dB, SSIM: 0.931, capacity: 0.773 bpp) Fig shows the performance comparison results of the proposed, Hu et al.’s [4], and Tai et al.’s algorithms [5] The PSNR results in Fig 5(a) reveal that the capacity versus distortion performance of the proposed algorithm is comparable to the conventional ones at the low capacity region and the proposed algorithm exhibits slightly improved performance at the high capacity region In addition, the performance is degraded when is used This is because a the causal window of size simple average prediction in (1) does not perform well for large window sizes A more improved performance is expected by using a higher order prediction or changing the window size adaptively The SSIM comparison results in Fig 5(b) more clearly show that the proposed algorithm outperforms the conventional methods The subjective visual quality evaluation is also performed in Fig To facilitate comparison, the magnified regions of the original and two watermarked images obtained by Tai et al.’s and the proposed algorithms are shown At the similar watermark capacity, we can see that the proposed algorithm provides higher quality embedded images without producing annoying artifacts Since the proposed algorithm produces perceptually improved watermarked images, a public user who does not have a knowledge on the original image could not recognize the existence of the watermark Thus the proposed algorithm is In this letter, we have presented an improved histogram modification based reversible data hiding technique In the proposed algorithm, unlike the conventional reversible techniques, the HVS characteristics are extensively exploited to alleviate the distortion caused by data embedding The edge and JND values are estimated by using the causal window, and thus no additional overhead is required to be embedded By using the estimated values, the embedding level is adaptively adjusted for each pixel The experimental results demonstrated that this pixel level adaptive embedding method can provide the superior visual quality of the embedded images The proposed technique effectively exploited the well-known HVS characteristics for reversible image data embedding When the proposed algorithm is applied to reversible video data embedding, the video related HVS characteristics such as motion blur and motion sharpening can be additionally considered to produce perceptually pleasant video sequences REFERENCES [1] J Fridrich, M Goljan, and R Du, “Lossless data embedding-new paradigm in digital watermarking,” EURASIP J Appl Signal Process., vol 2002, no 2, pp 185–196, Feb 2002 [2] J Tian, “Reversible data embedding using a difference expansion,” IEEE Trans Circuits Syst Video Technol., vol 13, pp 890–896, Aug 2003 [3] S Weng, Y Zhao, J.-S Pan, and R Ni, “Reversible watermarking based on invariability and adjustment on pixel pairs,” IEEE Signal Process Lett., vol 15, pp 721–724, Nov 2008 [4] Y Hu, H.-K Lee, and J Li, “DE-based reversible data hiding with improved overflow location map,” IEEE Trans Circuits Syst Video Technol., vol 19, pp 250–260, Feb 2009 [5] W.-L Tai, C.-M Yeh, and C.-C Chang, “Reversible data hiding based on histogram modification of pixel differences,” IEEE Trans Circuits Syst Video Technol., vol 19, pp 906–910, Nov 2009 [6] Z Ni, Y Q Shi, N Ansari, and W Su, “Reversible data hiding,” IEEE Trans Circuits Syst Video Technol., vol 16, pp 354–362, Mar 2006 [7] W Lin, L Dong, and P Xue, “Visual distortion gauge based on discrimination of noticeable contrast changes,” IEEE Trans Circuits Syst Video Technol., vol 15, pp 900–909, Jul 2005 [8] I Höntsch and L Karam, “Adaptive image coding with perceptual distortion control,” IEEE Trans Image Process., vol 11, no 3, pp 213–222, Mar 2002 [9] A Polesel, G Ramponi, and V J Mathews, “Image enhancement via adaptive unsharp masking,” IEEE Trans Image Process., vol 9, no 3, pp 505–510, Mar 2000 [10] CVG-USR Image Database, [Online] Available: http://decsai.ugr.es/ cvg/dbimagenes [11] Z Wang, A C Bovik, H R Sheikh, and E P Simoncelli, “Image quality assessment: From error visibility to structural similarity,” IEEE Trans Image Process., vol 13, no 4, pp 600–612, Apr 2004 ... luminance, and in (4), only the pixels in the causal window are used because only these pixels are available at the data extraction stage due to the raster scan order processing Actual data embedding... FEBRUARY 2011 suitable to the conventional applications of the reversible data hiding, such as art, medical, and military imaging In addition, note that the proposed algorithm produces the embedded... original image in some applications, where the sharp image details are preferred In such applications, the proposed algorithm can be used to perform the image enhancement and reversible data hiding