Вы находитесь на странице: 1из 10


Xiaolong Li, Weiming Zhang, Xinlu Gui, and Bin Yang

n this paper, based on two-dimensional differenceodication, a novel reversible data hiding (RDH) oposed by using difference-pair-mapping (DPM). idering each pixel-pair and its context, a sequence pairs of difference values is computed. Then, a nal difference-histogram is generated by counting of the resulting difference-pairs. Finally, reversible ing is implemented according to a specically deHere, the DPM is an injective mapping dened on rs. It is a natural extension of expansion embedding techniques used in current histogram-based RDH the proposed approach, compared with the convenmensional difference-histogram and one-dimensional ror-histogram-based RDH methods, the image ren be better exploited and an improved embedding is achieved. Moreover, a pixel-pair-selection strategy ed to priorly use the pixel-pairs located in smooth s to embed data. This can further enhance the emormance. Experimental results demonstrate that the eme outperforms some state-of-the-art RDH works.

msDifference-pair-mapping (DPM), histogram reversible data hiding (RDH), two-dimensional togram.


SIBLE data hiding (RDH) aims to embed seessage into a cover image by slightly modifying ues, and, unlike conventional data hiding, the message as well as the cover image should be ecovered from the marked content [1][3]. RDH type of information hiding and its feasibility is o the lossless compressibility of natural images. lity in RDH is quite desirable and helpful in some lications such as medical image processing [4],

ceived November 23, 2012; revised February 27, 2013 and accepted April 24, 2013. Date of publication May 01, 2013; version May 29, 2013. The work of W. Zhang was supported f Science and Technology on Information Assurance LaboraKJ-13-02. The associate editor coordinating the review of this

[5], multimedia archive management [6], image [7], and video error-concealment coding [8], etc. G performance of a RDH scheme is evaluated by the tortion behavior. For a required embedding capa obtain a good marked image quality, one expects embedding distortion as much as possible. Many RDH methods have been proposed the methods based on lossless compression [9] ence expansion [12][15], histogram modicatio prediction-error expansion [20][25], and integ [26][30], etc. Among them, the histogram-base attracted much attention. The histogram-based me the histogram in such a way that certain bins a create vacant space while some other bins are uti data by lling the vacant space. This type of meth control the embedding distortion and provide a su The rst histogram-based RDH method is the o by Ni et al. in [16]. This method uses peak a points of the pixel-intensity-histogram to em changes each pixel value at most by 1, and thus a image quality can be obtained. However, its EC and this method does not work well if the cov a at histogram. To facilitate it, Lee et al. [17] utilize the difference-histogram instead. This nove ploits the correlation among neighboring pixels an larger payload with reduced distortion compare al.s. Moreover, we will see later (see Section II et al.s method can be in fact implemented, in a way, by modifying the two-dimensional pixeltogram according to a pixel-pair-mapping (PP an injective mapping dened on pixel-pairs. In t superiority of Lee et al.s method over Ni et al.s in another viewpoint. Afterwards, Fallahpour [18 a method by modifying the histogram of pre Like difference-histogram, the prediction-erroralso Laplacian-like and sharply distributed whic an excellent embedding performance. Instead o the correlation of two adjacent pixels in Lee et Fallahpours method can exploit the local cor

fference value of a pixel-pair) to get a scalar seThen, a one-dimensional histogram (e.g., differstogram) is generated by counting the frequency of lting sequence. ram modication) Finally, embed data into the mage by modifying the histogram. In most cases, ogram bins with high frequencies are expanded data while some others are shifted to ensure the ility. step, the complex local image correlation is simne-dimensional statistic. Clearly, by this simplimage redundancy cannot be fully exploited and it utes to the second step since a one-dimensional hisy to deal with. Based on this consideration, instead nsional histogram used in current RDH methods exploit the image redundancy, we propose in this l RDH scheme by using a two-dimensional differam. oposed method, by considering a pixel-pair and its cal image region is projected to a two-dimensional tain a sequence consisting of difference-pairs. o-dimensional difference-histogram is generated the difference-pairs. Finally, reversible data is implemented according to a specically derence-pair-mapping (DPM). Here, the DPM is mapping dened on difference-pairs, and it is a nsion of expansion embedding and shifting techin current histogram-based methods. By using the onal difference-histogram and this specic DPM, ith the conventional one-dimensional histogram ds, more pixels are used for carrying data while of shifted pixels is reduced as well, and thus an mbedding performance is achieved. In addition, he embedding-position-selection techniques introvious works [24], [31], [34], a pixel-pair-selection dopted in our method to priorly use the pixel-pairs mooth image regions to embed data. This may nce the embedding performance. Experimental onstrate that the proposed method outperforms f-the-art works. of the paper is organized as follows. The related iey introduced in Section II. Section III presents RDH scheme in details. The comparisons with the e shown in Section IV. Section V concludes this

ding, the difference image cover image as

is rst computed fo

is a pixel-pair cons where consecutive pixels. Then the marked image ca as if if if if if

where is a data bit to be embedded. N rst pixel in the pair keeps unchanged in th procedure, i.e., one simply takes Accordingly, from a marked image , the em bit can be extracted as if if and the original pixel value can be recovered as if if if

where is t value computed from the marked image. In this method, the bins 1 and are utilized embedding and other bins (except bin 0) for shift forms some classical RDH methods such as [12] In 2009, another difference-histogram-based was proposed by Tai et al. [15]. This method co extend the embedding space by utilizing a binary (BTS). Particularly, unlike Lee et al.s, the pixel-p ence calculation in this method can be overlappe data extraction procedure, the rst pixel of the pair covered when processing this pixel-pair. In this ov and by utilizing BTS, Tai et al.s method perform can provide a higher EC than that of Lee et al.s. Recently, Tai et al.s method is improved by H utilizing a dual binary tree (DBT), a better pixel

Fig. 1. PPM for illustrating the data embedding procedure for (a) Lee et al.s method and (b) its improvement.

improve the image quality and EC of Tai et al.s te-of-the-art works such as [37] and [38]. mentioning that some recent works [18][21] can ed as Tai et al.s improvement. In [18][21], the lue is replaced by the prediction-error computed r local image region so that the redundancy can be ed. d is motivated by the aforementioned works and ibed in details in next section. III. PROPOSED RDH SCHEME

the proposed RDH scheme is based on modicaimensional difference-histogram by constructing a s an injective mapping dened on difference-pairs. on is organized as follows. First, to better exa, we introduce Lee et al.s method from a PPM Section III-A. We show that this difference-hisd method can be implemented by modifying the onal pixel-intensity-histogram according to a PPM injective mapping dened on pixel-pairs. Then II-B, we extend the idea of PPM to DPM and DPM-based scheme. In this way, by extending difgram (or, in an equivalent way, two-dimensional y-histogram) to two-dimensional difference-hisroposed scheme extends current histogram-based

parts as black points and blue points, each black po to a blue one (indicated by a green arrow) and ea is mapped to another blue point. Here, each point r value of a pixel-pair, and the black points are use sion embedding while the blue ones for shifting. this PPM, for a cover pixel-pair , its marked determined in the following way: 1) if (i.e., is a red point), the m pair is taken as itself. 2) if or (i.e., is a a) if the to-be-embedded data bit pixel-pair is taken as itself. b) if the to-be-embedded data bit pixel-pair is taken as its associate blue p 3) if or (i.e., is the marked pixel-pair is taken as its associate The corresponding data extraction and imag process can also be demonstrated according to th it is an injection, i.e., each point has at most one trivial description is omitted. From the PPM viewpoint, Lee et al.s differenc based method is actually implemented by modify dimensional pixel-intensity-histogram. Lee et al.s method only modies the second pix Thus two modication directions, up and down, a

Fig. 3. Context of , where the location of pixel is are used to compute the GAP predictor for , and all ten pixe are used to compute the noisy-level. has four modication direcalso has four modication is a prediction of .

modifying either or by 1, rresponding difference-pair e , , and

pixel-pair predictor:

, compute the prediction of to ge

xel-pairs (black points) are utilized for expansion and the number of shifted pixel-pairs (blue ones) is ell. Here in Fig. 1(b), the parameters and can y selected by maximizing EC. w PPM, more data bits can be embedded without he marked image quality. Taking the 512 512 cale Lena image for an example, compared with original method, EC is increased from 24,000 bits s and PSNR is slightly improved from 52.12 dB to he advantage of the new PPM lies in the exploitamodication directions. RDH

if if if if if if if where Fig. 3),

, we propose to compute two differand to form a two-difference-histogram of , where is a prewhich will be claried later. Inspired by the aforeew PPM, we will modify either or by 1. In , since has four modication directions, the air also has four modication directions: , or )). For example, by modifying to , the modction to is up and the corresponding modction to is upper-left, since changes d changes to . Based on these four modctions, we will introduce a new RDH scheme by DPM. of related works play an important role in our extend the idea of two-dimensional pixel-intenm (or, in an equivalent sense, one-dimensional stogram) of Lee et al. to two-dimensional difgram. Besides, for each , we compute the of based on the context of for an accurate Here, the gradient-adjusted-prediction (GAP) will


are neighboring pixels an represent the vertical a gradients, and . should be rounded to its nearest integer if it is n Then, compute the noisy-level of denoted using its ten neighboring pixels (se as

where represents the context of conta pixels and stands for the gradient operator. H crete image, the noisy-level is computed by su vertical and horizontal differences of every two pixels in , and it is less than or equal to Clearly, a pixel-pair located in smooth regions small noisy-level. Finally, for each pixel-pair wi less than a threshold , compute the difference and implement data embedding according to the in Fig. 4. Here, as the same as the case of PPM, each blac DPM is mapped to a blue one and it will be used one data bit, while each blue point is mapped to one and it is simply shifted to ensue the reversibil


M for illustrating the proposed data embedding procedure.

M will be experimentally demonstrated in the next

some explanations for the nal step of our clarify how DPM works. For example, when and , the corresponding differwhich is a black point in Fig. 4. In pixel-pair will be expanded to carry one data bit: be-embedded data bit , the marked pixel-pair as itself. o-be-embedded data bit , the marked dif-

. For general cases, as a scription of the DPM shown in Fig. 4, the marked v in Table I. Notice that the pixel-pair scanning order in ou tion procedure is inverse to that of embedding. B when processing a pixel-pair in data extraction, it already been recovered. Thus the same predictio level used by encoder can be obtained by decode and the injectivity of DPM guarantee the revers scheme. The corresponding data extraction and covery process are listed in Table II. In the table, and are difference values compu marked pixel-pair , where is the sam used in data embedding by encoder. C. Evaluation for DPM

As we have known (see Section II-C of [24] for the histogram-based RDH, if the maximum m pixel values is 1 in data embedding, the expected modication (in -norm) to cover image is

DPM-based scheme. Clearly, for our method, acg. 4, the ratio of expanded pixels can be formulated


means the frequency of . shows the comparisons of the ration of expanded r method with different pixel-pair-selection threshal.s method [17] and its improved one presented in A, Hong et al.s MED-based method [19], and FalAP-based method [18]. Except Barbara, all the imour experiments are downloaded from the USCe1. For our method, when the pixel-pair-selection s taken as its maximum, i.e., the sum of maximum f all pixel-pairs and 1 (9)

Fig. 5. Performance comparisons between the proposed me method [17] and its improved one presented in Section IIIMED-based method [19], and Fallahpours GAP-based metho Lena.

t the pixel-pair-selection strategy is disabled and s are used for data embedding. According to this n see that our method is better than those prior arts ys has a larger ratio. The superiority of the two-difference-histogram and this specically designed veried. On the other hand, for our method, the ras larger with smaller . This demonstrates the adixel-pair-selection strategy. Finally, as a complee III, Fig. 5 shows performance comparisons with tioned works. Our scheme performs well and it sigtperforms the prior arts. ed data embedding and extraction procedures inreatment of overow/underow will be given in section. More experimental results of comparisons -the-art works will be reported in Section IV.

by using LSB replacement, embed the auxiliary and the compressed location map into . The embedding procedure is described as bellow stepStep-1: Except the last two columns and last tw left to right and top to bottom, divide the cove nonoverlapped pixel-pairs. Then we dene a noted to solve the overow/underow problem the maximum modication to each pixel is 1 in o we only need to deal with the pixels with extrem 255. Specically, for the -th pixel-pair , if , we dene . Otherwise, we t Consequently, is a binary sequence of length lessly compress the location map using arithmeti note the length of the compressed location map a pixel-pair with , compute the predi the noisy-level according to (5) and(6) Step-2: The secret message will be embedded i cording to the location map and a pixel-pair-selec , i.e., successively, for each , con pixel-pair:

Fig. 6. Six test images. From left to right: Lena, Baboon, Barbara, Airplane (F-16), Peppers, and Fishing boat.

e auxiliary information and the compressed locabe embedded in this step for blind decoding. This several parts: cord LSB of the rst image o obtain a binary sequence , where is the unction. into the remaining pixel-pairs (i.e., the mbed irs with index to ) using the same method in After embedded, we denote dex of the last data-carrying pixel-pair. replace LSB of the rst pixels ollowing auxiliary information and the compressed map dened in Step-1 to generate the marked

according to Table II; otherwise, there is no da and the pixel-pair is just recovered as itself. It should be mentioned that for the -th pixel-pair w it is unchanged in embedding procedure, and thus can be recovered as itself. Step-3: Replace LSB of the rst pixels by the sequence extracted in the step abo the same method as Step-2 to extract the embed from the rst pixel-pairs, and meanwhile to re tion for these pixel-pairs. Finally, the embedded m tracted and the original image is recovered. IV. EXPERIMENTAL RESULTS

el-pair-selection threshold: (12 bits), ex of the last embedded pixel-pair in Step-2: bits), gth of the compressed location map: ( s), ex of the last embedded pixel-pair: ( s). that, in the above embedding procedure, to priorly ixel-pairs, the threshold is taken as the smallest ger such that the message can be embedded. This n be determined iteratively. sponding data extraction and image restoration ummarized as follows. ad LSB of the rst pixels of e to determine the values of , , and . e next LSB to determine the compressed location generate the location map by decompressing ed location map. this step, we extract the LSB sequence dened 3 of data embedding. First, as the same as data emept the last two columns and last two rows, divide mage into nonoverlapped pixel-pairs. Then, in a , from the -th to -th pixel-pair, extract and restore the cover image as follows. Suppose he to-be-processed pixel-pair is :

Six 512 512 sized gray-scale images incl Baboon, Barbara, Airplane (F-16), Peppers, and (see Fig. 6) are used in our experiment. Fig. performance comparisons between the proposed following state-of-the-art works: Lee et al. [17] improved one presented in Section III-A), Fall Hong et al. [19], Sachnev et al. [31], Li et al. [24 and Hong [33]. For our method, we vary EC from its maximum with a step of 1,000 bits. Here, 5,00 means an embedding rate (ER) of 0.019 bits per p According to this gure, compared with Lee et and its improvement, one can see that our super nicant. It experimentally demonstrates that the scheme can provide a much better performance th Moreover, compared with the methods of Fa Hong et al., our superiority is also signicant. The are based on the prediction-error-histogram w predictors. Although these methods may explo redundancy for a larger pixel context and perform Lee et al.s in most cases, our scheme can improv creasing PSNR by 16 dB. Our advantage lies in t of two-dimensional difference-histogram and pix tion strategy. Sachnev et al.s method is based on the predict togram incorporating with an embedding-posit strategy similar to ours. A sorting technique is

this paper, and we will investigate this issue in our V. CONCLUSION

TABLE V F PSNR (IN dB) BETWEEN OUR METHOD AND FOUR METHODS V et al. [31], LI et al. [24], HONG [34], AND HONG [33], FOR AN EC OF 20,000 BITS (ER 0.076 BPP)

and V, our method generally achieves a larger Sachnev et al.s. Our average gains are 1.44 dB 10,000 bits, and 1.2 dB for an EC of 20,000 bits. he result for Baboon is not presented in Table V hod cannot embed such a payload into this image. nd Hongs methods [24], [33], [34] are also based ction-error-histogram incorporating with an emtion-selection strategy. According to Fig. 7, as se methods perform similarly to Sachnev et al.s better than ours only for some images when our es its maximum. In most cases, our method can ger PSNR. Referring to Tables IV and V, one can method improves Li et al. and Hongs by increasing east 1 dB in average. that, comparing our method with [24], [31], [33], n advantage is the utilization of two-dimensional stogram. iscuss the size of the compressed location map. n our method, the maximum modication to each is at most 1 in value, so the overow/underow y only occur for the pixel-pair where . For the six test images Lena, Baa, Airplane (F-16), Peppers, and Fishing boat, exwo columns and last two rows which are not used dding, the amount of pixel-pairs satisfying or is 0, 0, 0, 0, 2, and 9, respectively. st cases, there are only a few 1 in the location

In this paper, we presented a novel RDH sche a two-dimensional difference-histogram according cally designed DPM. In addition, a pixel-pair-sele is also proposed to further enhance the embedding This work is the rst attempt to employ higher histogram to design RDH. Compared with the p troduced one-dimensional histogram based meth proach can exploit the image redundancy better an improved performance. However, since only a pixel-pair is allowed to be modied by 1 in v is low. This issue should be investigated in the f over, utilizing more suitable two-dimensional histo signing more meaningful DPM (e.g., in an imag way) to achieve the best embedding performance i able problem. REFERENCES

[1] Y. Q. Shi, Reversible data hiding , in Proc. IWDW, pp. 112, ser. Springer LNCS. [2] Y. Q. Shi, Z. Ni, D. Zou, C. Liang, and G. Xuan, Loss fundamentals, algorithms and applications, in Pro 2004, vol. 2, pp. 3336. [3] R. Caldelli, F. Filippini, and R. Becarelli, Reversib techniques: An overview and a classication, EURA rity, vol. 2010, 2010, Article 134546. [4] G. Coatrieux, C. L. Guillou, J. M. Cauvin, and C. Ro watermarking for knowledge digest embedding and r in medical images, IEEE Trans. Inf. Technol. Biomed pp. 158165, Mar. 2009. [5] M. Fontani, A. D. Rosa, R. Caldelli, F. Filippini, A. P salvo, Reversible watermarking for image integrity v erarchical pacs, in Proc. 12th ACM Workshop on Mu curity, 2010, pp. 161168. [6] S. Lee, C. D. Yoo, and T. Kalker, Reversible imag based on integer-to-integer wavelet transform, IE Forensics Security, vol. 2, no. 3, pp. 321330, Sep. 2 [7] R. Li, O. C. Au, C. K. M. Yuk, S. Yip, and T. Chan, trans-coding using reversible data hiding, in Proc. IE pp. 12731276. [8] K.-L. Chung, Y.-H. Huang, P.-C. Chang, and H.-Y. L data hiding-based approach for intra-frame error H.264/AVC, IEEE Trans. Circuits Syst. Video Techn 11, pp. 16431647, Nov. 2010. [9] J. Fridrich, M. Goljan, and R. Du, Lossless data e paradigm in digital watermarking, EURASIP J. Appl. vol. 2002, no. 2, pp. 185196, Feb. 2002. [10] M. U. Celik, G. Sharma, A. M. Tekalp, and E. Saber eralized-LSB data embedding, IEEE Trans. Image P no. 2, pp. 253266, Feb. 2005.

n watermarking, in Proc. IEEE ICME, 2006, pp. 13211324. lahpour, Reversible image data hiding based on gradient d prediction, IEICE Electron. Express, vol. 5, no. 20, pp. 6, Oct. 2008. ng, T. S. Chen, and C. W. Shiu, Reversible data hiding for uality images using modication of prediction errors, J. Syst. re, vol. 82, no. 11, pp. 18331842, Nov. 2009. Thodi and J. J. Rodriguez, Expansion embedding techniques for ble watermarking, IEEE Trans. Image Process., vol. 16, no. 3, 730, Mar. 2007. H. K. Lee, and J. Li, DE-based reversible data hiding with ed overow location map, IEEE Trans. Circuits Syst. Video l., vol. 19, no. 2, pp. 250260, Feb. 2009. uc, Improved embedding for prediction-based reversible waking, IEEE Trans. Inf. Forensics Security, vol. 6, no. 3, pp. 2, Sep. 2011. , L. An, Y. Yuan, D. Tao, and X. Li, Lossless data embedding eneralized statistical quantity histogram, IEEE Trans. Circuits deo Technol., vol. 21, no. 8, pp. 10611070, Aug. 2011. B. Yang, and T. Zeng, Efcient reversible watermarking based ptive prediction-error expansion and pixel selection, IEEE Image Process., vol. 20, no. 12, pp. 35243533, Dec. 2011. Wu and J. Huang, Reversible image watermarking on prediction by efcient histogram modication, Signal Process., vol. 92, pp. 30003009, Dec. 2012. uc and J. M. Chassery, Very fast watermarking by reversible t mapping, IEEE Signal Process. Lett., vol. 14, no. 4, pp. 8, Apr. 2007. g, Y. Zhao, J. S. Pan, and R. Ni, Reversible watermarking on invariability and adjustment on pixel pairs, IEEE Signal s. Lett., vol. 15, no. , pp. 721724, 2008. ng, X. Li, B. Yang, and Z. Guo, Efcient generalized integer rm for reversible watermarking, IEEE Signal Process. Lett., no. 6, pp. 567570, Jun. 2010. , X. Li, and B. Yang, Adaptive reversible data hiding scheme n integer transform, Signal Process., vol. 92, no. 1, pp. 5462, 12. tuc, Low distortion transform for reversible watermarking, rans. Image Process., vol. 21, no. 1, pp. 412417, Jan. 2012. hnev, H. J. Kim, J. Nam, S. Suresh, and Y. Q. Shi, Reversible arking algorithm using sorting and prediction, IEEE Trans. s Syst. Video Technol., vol. 19, no. 7, pp. 989999, Jul. 2009. Z. Chen, M. Chen, X. Zeng, and Z. Xiong, Reversible image arking using interpolation technique, IEEE Trans. Inf. Forencurity, vol. 5, no. 1, pp. 187193, Mar. 2010. ng, An efcient prediction-and-shifting embedding technique h quality reversible data hiding, EURASIP J. Adv. Signal s., vol. 2010, 2010. ng, Adaptive reversible data hiding method based on error enntrol and histogram shifting, Opt. Commun., vol. 285, no. 2, 108, 2012. g, X. Li, and B. Yang, High capacity reversible image waterg based on integer transform, in Proc. IEEE ICIP, 2010, pp. 0. Weinberger, G. Seroussi, and G. Sapiro, The LOCO-I lossless compression algorithm: Principles and standardization into s, IEEE Trans. Image Process., vol. 9, no. 8, pp. 13091324, 000.

Xiaolong Li received the B.S. deg University, Beijing, China, the M Ecole Polytechnique, Palaiseau the Ph.D. degree in mathematics Cachan, Cachan, France, in 1999, respectively. Before joining Peking University he worked as a postdoctoral fellow versity from 2007 to 2009. His rese image processing and information h

Weiming Zhang received the M degrees in 2002 and 2005, respec Zhengzhou Information Science Institute, Zhengzhou, China. Currently, he is an associate pr School of Information Science a University of Science and Techn His research interests include mul information hiding, and privacy pro

Xinlu Gui received the B.S. deg University, Beijing, China, in 2012 working toward the Master degr science at Peking University, Bei research interests are image process tion hiding.

Bin Yang received the B.S. and M.S puter science from Peking Universit in 1991 and 1994, respectively. Currently, he is a Professor with Computer Science and Technology sity. His research interests are imag information hiding.