3108

IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 21, NO. 6, JUNE 2012

REFERENCES

[1] E. J. Cands, J. Romberg, and T. Tao, Robust uncertainty principles: Exact signal reconstruction from highly incomplete frequency information, IEEE Trans. Inf. Theory, vol. 52, no. 2, pp. 489509, Feb. 2006. [2] D. L. Donoho, Compressed sensing, IEEE Trans. Inf. Theory, vol. 52, no. 4, pp. 12891306, Apr. 2006. [3] R. G. Baraniuk, V. Cevher, M. Duarte, and C. Hegde, Model-based compressive sensing, IEEE Trans. Inf. Theory, vol. 56, no. 4, pp. 19822001, Apr. 2010. [4] L. He and L. Carin, Exploiting structure in wavelet-based Bayesian compressive sensing, IEEE Trans. Signal Process., vol. 57, no. 9, pp. 34883497, Sep. 2009. [5] J. Huang, D. Metaxas, and T. Zhang, Learning with structured sparsity, in ACM Int. Conf. Proc. Ser., 2009, vol. 382, pp. 417424. [6] S. Mun and J. E. Fowler, Block compressed sensing of images using directional transforms, in Proc. IEEE ICIP, 2009, pp. 30213024. [7] X. Wu, X. Zhang, and J. Wang, Model-guided adaptive recovery of compressive sensing, in Proc. Data Compression Conf., Snowbird, UT, 2009, pp. 123132. [8] P. J. Garrigues, Sparse coding models of natural images: Algorithms for efcient inference and learning of higher-order structure, Ph.D. Scalable Coding of Encrypted Images dissertation, Univ. California, Berkeley, CA, 2009. [9] Y. Kim, M. S. Nadar, and A. Bilgin, Exploiting wavelet-domain deXinpeng Zhang, Member, IEEE, Guorui Feng, Yanli Ren, and pendencies in compressed sensing, in Proc. Data Compression Conf., Zhenxing Qian Snowbird, UT, 2010, p. 536. [10] S. G. Mallat, A Wavelet Tour of Signal Processing: The Sparse Way. Amsterdam, The Netherlands: Elsevier, 2009. [11] M. J. Wainwright and E. P. Simoncelli, Scale mixtures of Gaussians and the statistics of natural images, Adv. Neural Inf. Process. Syst., AbstractThis paper proposes a novel scheme of scalable coding for vol. 12, no. 1, pp. 855861, 2000. encrypted images. In the encryption phase, the original pixel values are [12] S. G. Chang, B. Yu, and M. Vetterli, Spatially adaptive wavelet threshmasked by a modulo-256 addition with pseudorandom numbers that are olding with context modeling for image denoising, IEEE Trans. Image derived from a secret key. After decomposing the encrypted data into a Process., vol. 9, no. 9, pp. 15221531, Sep. 2000. downsampled subimage and several data sets with a multiple-resolution [13] J. M. Shapiro, Embedded image coding using zerotrees of wavelet construction, an encoder quantizes the subimage and the Hadamard coefcients, IEEE Trans. Signal Process., vol. 41, no. 12, pp. coefcients of each data set to reduce the data amount. Then, the data of 34453462, Dec. 1993. quantized subimage and coefcients are regarded as a set of bitstreams. http://ieeexploreprojects.blogspot.com [14] A. Said and W. A. Pearlman, A new, fast, and efcient image codec At the receiver side, while a subimage is decrypted to provide the rough based on set partitioning in hierarchical trees, IEEE Trans. Circuits information of the original content, the quantized coefcients can be used Syst. Video Technol., vol. 6, no. 3, pp. 243250, Jun. 1996. to reconstruct the detailed content with an iteratively updating procedure. [15] D. S. Taubman and M. W. Marcellin, JPEG2000: Image Compression Because of the hierarchical coding mechanism, the principal original Fundamentals, Standards, and Practice. Boston, MA: Kluwer, 2002. content with higher resolution can be reconstructed when more bitstreams [16] Y. M. Lu and M. N. Do, Sampling signals from a union of subspaces, are received. IEEE Signal Process. Mag., vol. 25, no. 2, pp. 4147, Mar. 2008. Index TermsHadamard transform, image compression, image encryp[17] T. Blumensath and M. E. Davies, Sampling theorems for signals from the union of nite-dimensional linear subspaces, IEEE Trans. tion, scalable coding. Inf. Theory, vol. 55, no. 4, pp. 18721882, Apr. 2009. [18] C. Hegde, M. F. Duarte, and V. Cevher, Compressive sensing recovery of spike trains using a structured sparsity model, presented I. INTRODUCTION at the Signal Processing Adaptive Sparse Structured Representations Conf., Saint-Malo, France, 2009, Paper EPFL-CONF-151471. In recent years, encrypted signal processing has attracted con[19] M. N. Do and C. N. H. La, Tree-based majorize-maximize algorithm siderable research interests [1]. The discrete Fourier transform and for compressed sensing with sparse-tree prior, Proc. IEEE Int. Workadaptive ltering can be implemented in the encrypted domain based shop on Computational Advances in Multi-Sensor Adaptive Processing on the homomorphic properties of a cryptosystem [2], [3], and a (CAMPSAP 2007), pp. 129132, 2007. composite signal representation method can be used to reduce the size [20] E. J. Cands, M. B. Wakin, and S. P. Boyd, Enhancing sparsity by reweighted l minimization, J. Fourier Anal. Appl., vol. 14, no. 5, pp. of encrypted data and computation complexity [4]. In joint encryption 877905, 2008. and data hiding, a part of signicant data of a plain signal is encrypted [21] I. Daubechies, R. DeVore, M. Fornasier, and C. S. Gntrk, Iteratively reweighted least squares minimization for sparse recovery, Commun. for content protection, and the remaining data are used to carry the additional message for copyright protection [5], [6]. With some Pure Appl. Math., vol. 63, no. 1, pp. 138, Jan. 2010. [22] T. Blumensath and M. E. Davies, Iterative thresholding for sparse approximations, J. Fourier Anal. Appl., vol. 14, no. 5/6, pp. 629654, Manuscript received July 26, 2011; revised October 29, 2011 and December 2008. [23] J. Portilla, V. Strela, M. J. Wainwright, and E. P. Simoncelli, Image 15, 2011; accepted January 26, 2012. Date of publication February 13, 2012; date of current version May 11, 2012. This work was supported in part by the denoising using scale mixtures of Gaussians in the wavelet domain, IEEE Trans. Image Process., vol. 12, no. 11, pp. 13381351, Nov. National Natural Science Foundation of China under Grant 61073190, Grant 61103181, and Grant 60832010, and in part by the Alexander von Humboldt 2003. Foundation. The associate editor coordinating the review of this manuscript and [24] Compressive Sensing Resources [Online]. Available: http://dsp.rice. approving it for publication was Dr. Anthony Vetro. edu/cs The authors are with School of Communication and Information Engineering, [25] R. Garg and R. Khandekar, Gradient descent with sparsication: An Shanghai University, Shanghai 200072, China (e-mail: xzhang@shu.edu.cn). iterative algorithm for sparse recovery with restricted isometry propColor versions of one or more of the gures in this paper are available online erty, in Proc. 26th Annu. Int. Conf. Mach. Learn., 2009, pp. 337344. at http://ieeexplore.ieee.org. [26] D. L. Donoho and I. M. Johnstone, Ideal spatial adaptation by wavelet shrinkage, Biometrika, vol. 81, no. 3, pp. 425455, Aug. 1994. Digital Object Identier 10.1109/TIP.2012.2187671 1057-7149/$31.00 2012 IEEE

[27] T. T. Do, T. D. Tran, and L. Gan, Fast compressive sampling with structurally random matrices, in Proc. IEEE ICASSP, 2008, pp. 33693372. [28] L. Gan, T. T. Do, and T. D. Tran, Fast compressive imaging using scrambled block Hadamard ensemble, in Proc. Eur. Signal Process. Conf. (EUSIPCO), Lausanne, Switzerland, 2008. [29] E. J. Cands, J. Romberg, and T. Tao, Stable signal recovery from incomplete and inaccurate measurements, Commun. Pure Appl. Math., vol. 59, no. 8, pp. 12071223, Aug. 2006. [30] l Magic Toolbox [Online]. Available: http://users.ece.gatech.edu/ ~justin/l1magic/ [31] NESTA toolbox [Online]. Available: http://acm.caltech.edu/~nesta/ [32] Sparsify toolbox [Online]. Available: http://www.personal. soton.ac.uk/tb1m08/ [33] USC-SIPI image database [Online]. Available: http://sipi.usc.edu/database/

IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 21, NO. 6, JUNE 2012

3109

buyerseller protocols [7], [8], the ngerprint data are embedded into and coefcients are regarded as a set of bitstreams. When having the an encrypted version of digital multimedia to ensure that the seller encoded bitstreams and the secret key, a decoder can rst obtain an cannot know the buyers watermarked version while the buyer cannot approximate image by decrypting the quantized subimage and then obtain the original product. reconstructing the detailed content using the quantized coefcients A number of works on compressing encrypted images have been also with the aid of spatial correlation in natural images. Because of the presented. When a sender encrypts an original image for privacy pro- hierarchical coding mechanism, the principal original content with tection, a channel provider without the knowledge of a cryptographic higher resolution can be reconstructed when more bitstreams are key and original content may tend to reduce the data amount due to the received. limited channel resource. In [9], the compression of encrypted data is investigated with the theory of source coding with side information at A. Image Encryption the decoder, and it is pointed out that the performance of compressing Assume that the original image is in an uncompressed format and encrypted data may be as good as that of compressing nonencrypted data in theory. Two practical approaches are also given in [9]. In the that the pixel values are within [0, 255], and denote the numbers of rst one, the original binary image is encrypted by adding a pseudo- rows and columns as 1 and 2 and the pixel number as ( = 1 2 2 ). Therefore, the bit amount of the original image is 8 . The random string, and the encrypted data are compressed by nding the syndromes of low-density parity-check (LDPC) channel code. In the content owner generates a pseudorandom bit sequence with a length second one, the original Gaussian sequence is encrypted by adding of 8 . Here, we assume the content owner and the decoder has the an independent identically distributed Gaussian sequence, and the en- same pseudorandom number generator (PRNG) and a shared secret key crypted data are quantized and compressed as the syndromes of trellis used as the seed of the PRNG. Then, the content owner divides the code. While Schonberg et al. [10] study the compression of encrypted pseudorandom bit sequence into pieces, each of which containing 8 data for memoryless and hidden Markov sources using LDPC codes, bits, and converts each piece as an integer number within [0, 255]. An Lazzeretti and Barni [11] present several lossless compression methods encrypted image is produced by a one-by-one addition modulo 256 as for encrypted gray and color images by employing LDPC codes into follows: various bit planes. In [12], the encrypted image is decomposed in a (0) ( ) = mod [ ( )+ ( ) 256] progressive manner, and the data in most signicant planes are com1   1 1   2 (1) pressed using rate-compatible punctured turbo codes. Based on local statistics of a low-resolution version of the image, the original content can be perfectly reconstructed. By extending the statistical models to where ( ) represents the gray values of pixels at positions ( ), video, some algorithms for compressing encrypted video are presented ( ) represents the pseudorandom numbers within [0, 255] generin [13]. In most of aforementioned schemes, the syndrome of channel ated by the PRNG, and (0) ( ) represents the encrypted pixel values. code is exploited to generate the compressedhttp://ieeexploreprojects.blogspot.com Lena and its encrypted version. data in a lossless manner. Fig. 1 gives an original image Furthermore, several methods for lossy compressing encrypted imClearly, the encrypted pixel values (0) ( ) are pseudorandom ages have been developed. In [14], a compressive sensing mechanism numbers since ( ) values are pseudorandom numbers. It is well is introduced to achieve the lossy compression of encrypted images, known that there is no probability polynomial time (PPT) algorithm to and a basis pursuit algorithm is used to enable joint decompression distinguish a pseudorandom number sequence and a random number and decryption. In [15], the original gray image is encrypted by pixel sequence until now. Therefore, any PPT adversary cannot distinguish permutation; then, the encrypted data are compressed by discarding the an encrypted pixel sequence and a random number sequence. That excessively rough and ne information of coefcients generated from is to say, the image encryption algorithm that we have proposed is orthogonal transform. When having the compressed data and the per- semantically secure against any PPT adversary. mutation way, a receiver can reconstruct the principal content of the original image by retrieving the values of coefcients. However, the B. Encrypted Image Encoding ratedistortion performance in [14] is low, and there is a leakage of Although an encoder does not know the secret key and the original statistical information in [15] since only the pixel positions are shufcontent, he can still compress the encrypted data as a set of bitstreams. ed and the pixel values are not masked in the encryption phase. This paper proposes a novel scheme of scalable coding for encrypted The detailed encoding procedure is as follows. First, the encoder decomposes the encrypted image into a series of gray images. Although there have been a lot of works on scalable coding of unencrypted images/videos [16], [17], the scalable coding subimages and data sets with a multiple-resolution construction. The (t+1) is generated by downsampling of encrypted data has not been reported. In the encryption phase of subimage at the ( + 1)th level the subimage at the th level as follows: the proposed scheme, the pixel values are completely concealed so

NN N

g i; j e i;j

p i;j

e i;j ;

i N;

j N

p i;j

i; j

g i; j

e i;j

g i; j

that an attacker cannot obtain any statistical information of an original image. Then, the encrypted data are decomposed into several parts, and each part is compressed as a bitstream. At the receiver side with the cryptographic key, the principal content with higher resolution can be reconstructed when more bitstreams are received. II. PROPOSED SCALABLE CODING SCHEME In the proposed scheme, a series of pseudorandom numbers derived from a secret key are used to encrypt the original pixel values. After decomposing the encrypted data into a subimage and several data sets with a multiple-resolution construction, an encoder quantizes the subimage and the Hadamard coefcients of each data set to effectively reduce the data amount. Then, the quantized subimage

g( +1) (i; j ) = g( ) (2i; 2j ); t = 0; 1; . . . ; T 0 1

t t

(2)

where (0) is just the encrypted image and is the number of decomposition levels. In addition, the encrypted pixels that belong to (t) but do not belong to (t+1) form data set (t+1) as follows:

Q( +1) = g( )(i; j )j mod (i; 2) = 1 or mod (j; 2) = 1 ;

t t

t = 0; 1; . . . ; T 0 1:

(3)

That means each (t) is decomposed into (t+1) and (t+1) , and the data amount of (t+1) is three times of that of (t+1) . After the multiple-level decomposition, the encrypted image is reorganized as (T ) (T ) (T 01) (1) . . ., and .

G Q

G ;Q ;Q

3110

IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 21, NO. 6, JUNE 2012

Fig. 1. (a) Original image Lena and (b) its encrypted version.

For the subimage step as follows:

G T , the encoder quantizes each value using a

( )

b i; j

( )=

g (T ) i; j

( ) 1

(4)

where 0 is a transpose of , is an L(t) 2 L(t) identity matrix, and (t) L(t) must be a multiple of 4. For each coefcient Ck l , calculate (t) Ck l ; (t) (t) (t) ck l ; kK ; lL (10) =M (t)

HI

( )=

mod ( ) 256 256

M

()

where the operator b1c takes an integer toward minus innity and

where (5)

1 = 256=M:

http://ieeexploreprojects.blogspot.com (t)

= round

p M= L(t)

(11)

Here, M is an integer shared by the encoder and the decoder, and its value will be discussed later. Clearly

0  b(i; j )  M 0 1: ( )

(6)

Then, the data of b i; j are converted into a bitstream, which is denoted as BG. The bit amount of BG is
NBG

For each data set (t) t ; ; ; T , the encoder permutes and divides encrypted pixels in it into K (t) groups, each of which conN= t . In this way, the L(t) taining L(t) pixels K (t) 2 L(t) pixels in the same group scatter in the entire image. The permutation way is shared by the encoder and the decoder, and the values of L(t)

Q ( = 1 2 ... )
(

= 4T 1 log
N

M:

(7)

(t) and round 1 nds the nearest integer. In (10), the remainder of Ck l (t) modulo 256 is quantized as integer ck l , and the quantization steps (t) are approximately proportional to square roots of L(t) . Then, ck l at different levels are converted into bitstreams, which are denoted as BS(t) . Since (t) (t)  ck l  M 0 (12) (t) and the number of ck l at the tth level is N= t , the bit amount of (t) BS is

()

()

() ()

0 ()

()

3 4

(t)

= 3 1 N 1 log 4t ...

M (t)

= 1 ; 2; . . . ; T :

(13)

= 3 4)

will be discussed later. Denote the encrypted pixels of the k th group (t) (t) (t) ; qk ; ; qk L (t)  k  K (t) , and perform the as qk Hadamard transform in each group as follows:

(1)

(2) . . .

) (1

Ck (t) Ck Ck
(t)

(t)

. . .

(1) (2)

=H1

qk (t) qk qk
(t)

(t)

L (t)

. . .

(1) (2)

The encoder transmits the bitstreams with an order of ; BS(1) g. If the channel bandwidth is limited, the latter bitstreams may be abandoned. A higher resolution image can be reconstructed when more bitstreams are obtained at the receiver side. Here, the total compression ratio RC , which is a ratio between the amount of the encoded data and the encrypted image data, is T T (t) NBG (t) 2M 2M RC N : (14) 1 t N N 1 T t=1 t=1
fBG; BS(T ) ; BS(T 01) ;

= 8 + 81

= log 4 + 3 8 8

log

(8) C. Image Reconstruction With the bitstreams and the secret key, a receiver can reconstruct the principal content of the original image, and the resolution of the reconstructed image is dependent on the number of received bitstreams. While BG provides the rough information of the original content, BS(t) can be used to reconstruct the detailed content with an iteratively updating procedure. The image reconstruction procedure is as follows.

L (t)

where is a L(t) 2 L(t) Hadamard matrix made up of meets implies the matrix

H H 1H=H1H =L t 1I
0 0

+1 or 01. That
(9)

( )

IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 21, NO. 6, JUNE 2012

3111

When having the bitstream BG, the decoder may obtain the values of b(i; j ) and decrypts them as a subimage, i.e.,
p
(T )

(i; j ) = mod b(i; j ) 1 1 0 e(2 1

i; 2


j ); 256 +
1


1 2

N1
2T

N2
2T

(15)

where e(2T 1 i; 2T 1 j ) are derived from the secret key. If the bitstreams BS(t) (  t  T ) are also received, an image with a size of N1 =2( 01) 2 N2 =2( 01) will be reconstructed. First, upsample the subimage p(T ) (i; j ) by factor 2(T 0 +1) to yield an N1 =2( 01) 2 N2 =2( 01) image as follows:
r (2

That means the transform of errors in the plain domain is equivalent to the transform of errors in the encrypted domain with the modular arithmetic. Denoting (t) ^ (t) rk (1) ^ Ck (1) (t) ^ (t) (2) Ck rk (2) ^ = 1 (22) . . . . . . (t) ^ (t) L (t) L (t) rk ^ Ck

we have
1Ck (l)
(t)


^ Ck (l) 0 Ck (l) mod 256:

(t)

(t)

(23)

T 0 +1

i; 2

T 0 +1

j) = p

(T )

(i; j ); 1


N1
2T

; 1

N2
2T

(16)

and estimate the values of other pixels according to the pixel values in (16) using a bilinear interpolation method. Then, the interpolated pixels are reorganized as data sets with multiple-resolution construction, and the data in each set are permuted and divided into a series of groups. Denote the interpolated pixel values of the k th group at the tth level as (t) (t) (t) rk (1); rk (2); . . . ; rk (L(t) ) (1  k  K (t) ;   t  T ) and their (t) (t) (t) corresponding original pixel values as pk (1); pk (2); . . . ; pk (L(t) ). The errors of interpolated values are
1pk (l) = pk (l) 0 rk (l); 1


(t) With the bitstreams BS(t) (  t  T ), the values of ck (l) can (t) be retrieved, which provide the information of Ck (l). Therefore, the receiver may use an iterative procedure to progressively improve the quality of the reconstructed image by updating the pixel values ac(t) cording to ck (l). The detailed procedure is as follows. (t) (t) (t) (t) ^ 1) For each group [rk (1); rk (2); . . . ; rk (L(t) )], calculate rk (l) ^ (t) (l) using (18) and (22). and Ck 2) Calculate

Dk (l) = mod ck (l) 1 1

~ Dk (l) =
(t) (t)

(t)

(t)

(t)

+1

(t)

^ =2 0 Ck (l); 256

(t)

(24) (25)

D k (l ); Dk (l) 0 256;

if dk (l) < 128 : if dk (l)  128

(t)

(t)

(t)

(t)

;1

(t)

; 

T:

(17)

(t) Dene the encrypted values of rk (l) as

follows: http://ieeexploreprojects.blogspot.com
rk (1)  (t) rk (2) 


~ Dk (l) are the differences between the values consistent with the (t) ^ (t) ~ (t) corresponding ck (l) and Ck (l). Then, considering Dk (l) as (t) an estimate of 1Ck (l), modify the pixel values of each group as
(t)

rk (l) = mod rk (l) + ek (l); 256 ; ^

1


(t)

(t)

(t)

(t)

; 1

(t)

; 

(18)

rk (L) 

(t)

. . .

rk (1) (t) rk (2) rk (L)

(t)

(t)

. . .

L (t)

~ Dk (1) ~ (t) Dk (2) ~ Dk

(t)

(t)

. . .

(26)

L (t)

(t) where ek (l) are pseudorandom numbers derived from the secret key (t) and corresponding to rk (l). Then

and enforce the modied pixel values into [0, 255] as follows:
rk (l) = ~
(t) (t) if rk (l) < 0  (t) rk (l); if 0  rk (l)  255   (t) 255 ; if rk (l) > 255 

1pk (l)

(t)

qk (l) 0 rk (l) mod 256: ^

(t)

(t)

(19)

(t)

(27)

We also dene
1Ck (2) 1Ck
(t)

1Ck (1)
(t)

(t)

. . .

1pk (2) 1pk

(t)

1pk (1)
(t)

(t)

3) Calculate the average energy of difference due to the modication as follows: (20)
D=

(t)

. . .

T K

(t)

where is a L(t) 2 L(t) Hadamard matrix made up of +1 or 01. Since only the addition and subtraction are involved in the Hadamard transform
1Ck (1)
(t) (t)

t= k=1 l=1 T

rk (l) 0 rk (l) ~
3N=4t

(t)

(t)

(28)

t= 

1Ck (2) 1Ck

(t)

qk (1)


(t) (t)

. . .

qk (2)

L (t)

qk

(t)

. . .

L (t)
(t) (t)

rk (2) ^ rk ^
(t)

rk (1) ^

. . .

mod 256:

(21)

L (t)

(t) If D is not less than a given threshold of 0.10, for each pixel rk (l), ~ after putting it back to the position in the image and regarding the av(t) erage value of its four neighbor pixels as its new value rk (l), go to Step 1. Otherwise, terminate the iteration, and output the image as a nal reconstructed result. In the iterative procedure, while the decrypted pixels p(T ) (i; j ) are (t) used to give an initial estimation of other pixels, the values of ck (l) in (t) bitstreams BS provide more detailed information to produce the nal reconstructed result with satisfactory quality. In Step 2, by estimating (t) (t) 1Ck (l) according to ck (l), the pixel values are modied to lower the reconstruction errors. If the image is uneven and L(t) is big, the (t) absolute value of actual 1Ck (l) may be more than 128 due to the

3112

IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 21, NO. 6, JUNE 2012

Fig. 2. Reconstructed Lena using fBS g, fBS ; BS g, fBS ; BS ; BS g, and fBS ; BS ; BS ; BS g. The values of PSNR in (a), (b), (c), and (d) when regarding the corresponding downsampled versions of original Lena as references are 38.4, 34, 37.1, and 38.4 dB. TABLE I COMPRESSION RATIOS, PSNR IN RECONSTRUCTED RESULTS AND ITERATION NUMBERS WITH DIFFERENT WHEN T ,L ,L , AND L WERE USED FOR LENA AND MAN

=3

=4

=8

= 24

http://ieeexploreprojects.blogspot.com

error accumulation in a group, so that ~ k ( ) in (25) may be not close (t) to 1 k ( ). To avoid this case, we let (t) decrease with a increasing since the spatial correlation in a subimage with lower resolution is weaker. For instance, (1) = 24, (2) = 8, and (3) = 4 for = 3. Furthermore, in Step 3, the value of each pixel is assigned as the average of its four neighbors to further approach its original value. Although the estimate of a certain pixel may be very different from its original value, the updating operation in Step 3 can effectively lower the error on the pixel since its neighbors are probably modied well. At last, we terminate the iterative procedure when the reconstruction quality is not improved further. Here, the small threshold of 0.10 ensures the convergence of iterative procedure.

C l

D l L
(t)

Two test images Lena and Man that are sized 512 2 512 were used as the original images in the experiment. We let = 3 and encoded = 24, (3) = 4, (2) = 8, and the encrypted images using (1) = 24 to produce the bitstreams BG, BS(3) , BS(2) , and BS(1) . In this case, the total compression ratio C = 0 318. Fig. 2 gives the reconstructed Lena using fBGg, fBG BS(3) g, fBG BS(3) BS(2) g and fBG BS(3) BS(2) BS(1) g, respectively. Reconstructed results with higher resolution were obtained when more bitstreams were used. When regarding the corresponding downsampled versions of original images as reference, the values of PSNR in reconstructed results are

III. EXPERIMENTAL RESULTS AND DISCUSSION

L ;

R ;

denoted as PSNRB , PSNR3 , PSNR2 , and PSNR1 . While the PSNR values of Lena are 38.4, 34, 37.1, and 38.4 dB, those of Man are 38.4, 31.9, 33.9, and 37.1 dB. In addition, the iterative updating procedure signicantly improved the reconstruction quality. For example, while PSNR in an interpolated 512 2 512, Lena is 23.9 dB; this value in the nal reconstructed image is 38.4 dB with a gain of 14.5 dB. Table I lists the compression ratios; the PSNR in reconstructed results and the numbers of iterations with respect to different when = 3, (3) = 4, (2) = 8, and (1) = 24 were used for images Lena and Man. All the encryption, encoding and reconstruction procedures were nished in several seconds by a personal computer. is larger, the compression ratio is higher, and the When the value of (t) reconstruction quality are better since k ( ) provide more detailed information. As there is less texture/edge content in Lena than Man, the quality of reconstructed Lena is better than that of Man. In addition, the larger (t) corresponds to the lower compression ratio and more (t) detailed k ( ). When we changed ( (3) (2) (1) ) from (4, 8, 12) to (4, 12, 32), the compression ratio decreased from 0.318 to 0.283, and the value of PSNR1 in reconstructed Lena and Man were 37.8 and 35.2 dB, respectively. Compared with the results in Table I, the new C -PSNR1 performance of Lena is better, whereas that of Man is worse. The reason is that Lena is smoother than Man. For Lena, the larger (t) was helpful to uniformly distribute the errors on pixels into (t) the Hadamard coefcients, and most of 1 k ( ) still fell into [0128,

c l

L c l

L ;L ;L

R L

C l

IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 21, NO. 6, JUNE 2012

3113

By discarding the encrypted data in the lowest bit planes, the method in [12] can be extended to achieve lossy compression. The performance of the extended method, which is also given in Fig. 4, is better than that of the proposed scheme. However, a decoder with higher computation complexity and the decoders feedback for sending rate of each bit plane are required in the method extended from [12]. That means the proposed scheme is more suitable for real-time decompression and some scenarios without feedback channel. IV. CONCLUSION This paper has proposed a novel scheme of scalable coding for encrypted images. The original image is encrypted by a modulo-256 addition with pseudorandom numbers, and the encoded bitstreams are made up of a quantized encrypted subimage and the quantized remainders of Hadamard coefcients. At the receiver side, while the subimage is decrypted to produce an approximate image, the quantized data of Hadamard coefcients can provide more detailed information for image reconstruction. Since the bitstreams are generated with a multiple-resolution construction, the principal content with higher resolution can be obtained when more bitstreams are received. The lossy compression and scalable coding for encrypted image with better performance deserves further investigation in the future. ACKNOWLEDGMENT The authors would like to thank the anonymous reviewers for their valuable comments.

Fig. 3. Performance of the proposed scheme with different T .

http://ieeexploreprojects.blogspot.com REFERENCES
[1] Z. Erkin, A. Piva, S. Katzenbeisser, R. L. Lagendijk, J. Shokrollahi, G. Neven, and M. Barni, Protection and retrieval of encrypted multimedia content: When cryptography meets signal processing, EURASIP J. Inf. Security, vol. 2007, pp. 120, Jan. 2007. [2] T. Bianchi, A. Piva, and M. Barni, On the implementation of the discrete Fourier transform in the encrypted domain, IEEE Trans. Inf. Forensics Security, vol. 4, no. 1, pp. 8697, Mar. 2009. [3] J. R. Troncoso-Pastoriza and F. Prez-Gonzlez, Secure adaptive ltering, IEEE Trans. Inf. Forensics Security, vol. 6, no. 2, pp. 469485, Jun. 2011. [4] T. Bianchi, A. Piva, and M. Barni, Composite signal representation for fast and storage-efcient processing of encrypted signals, IEEE Trans. Inf. Forensics Security, vol. 5, no. 1, pp. 180187, Mar. 2010. [5] S. Lian, Z. Liu, Z. Ren, and H. Wang, Commutative encryption and watermarking in video compression, IEEE Trans. Circuits Syst. Video Technol., vol. 17, no. 6, pp. 774778, Jun. 2007. [6] M. Cancellaro, F. Battisti, M. Carli, G. Boato, F. G. B. Natale, and A. Neri, A commutative digital image watermarking and encryption method in the tree structured Haar transform domain, Signal Process. Image Commun., vol. 26, no. 1, pp. 112, Jan. 2011. [7] N. Memon and P. W. Wong, A buyer-seller watermarking protocol, IEEE Trans. Image Process., vol. 10, no. 4, pp. 643649, Apr. 2001. [8] M. Kuribayashi and H. Tanaka, Fingerprinting protocol for images based on additive homomorphic property, IEEE Trans. Image Process., vol. 14, no. 12, pp. 21292139, Dec. 2005. [9] M. Johnson, P. Ishwar, V. M. Prabhakaran, D. Schonberg, and K. Ramchandran, On compressing encrypted data, IEEE Trans. Signal Process., vol. 52, no. 10, pp. 29923006, Oct. 2004. [10] D. Schonberg, S. C. Draper, and K. Ramchandran, On blind compression of encrypted correlated data approaching the source entropy rate, in Proc. 43rd Annu. Allerton Conf., Allerton, IL, 2005. [11] R. Lazzeretti and M. Barni, Lossless compression of encrypted greylevel and color images, in Proc. 16th EUSIPCO, Lausanne, Switzerland, Aug. 2008 [Online]. Available: http://www.eurasip.org/Proceedings/Eusipco/Eusipco2008/papers/1569105134.pdf [12] W. Liu, W. Zeng, L. Dong, and Q. Yao, Efcient compression of encrypted grayscale images, IEEE Trans. Signal Process., vol. 19, no. 4, pp. 10971102, Apr. 2010.

Fig. 4. Performance comparison of several compression methods.

128], so that the quality of reconstructed result was better. For Man, (t) the excessively large L(t) caused more Ck l with absolute values bigger than 128, leading to a lower reconstruction quality. Fig. 3 gives the RC -PSNR1 curves with different values of T . When an encrypted image is decomposed within more levels, more data are involved in quantization and compression; therefore, the RC -PSNR1 performance is better, and more iterations for image reconstruction are required. It is also shown that the performance improvement is not signicant when using a higher T more than 3. We also compare the proposed scheme with the previous methods and unencrypted JPEG compression in Fig. 4. Because it is difcult to completely remove the spatial data redundancy by the operations in the encrypted domain, the ratedistortion performance of the proposed scheme is signicantly lower than that of JPEG compression. On the other hand, the proposed scheme outperforms the method in [15]. With the method in [15], the original image is encrypted by pixel permutation, which implies an attacker without the knowledge of the secret key can know the original histogram from an encrypted image. In this proposed scheme, the original values of all pixels are encrypted by a modulo-256 addition with pseudorandom numbers, leading to semantic security. That means the attacker cannot obtain the original histogram from an encrypted image. In addition, the method in [15] does not support the function of scalable coding. Liu et al. [12] proposed a lossless compression method for encrypted images in a bit-plane based fashion.

()

3114

IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 21, NO. 6, JUNE 2012

and wide-swath imaging, calls for larger coding efciency at reduced encoding complexity. NASAs Solar TErrestrial RElations Observatory (STEREO) is providing groundbreaking images of the Sun using two space-based observatories.1 These images aim to reveal the processes in the solar surface (photosphere) through the transition region into the corona and provide the 3-D structure of coronal mass ejections (CMEs). A variety of image compression tools are currently used in deepspace missions, ranging from Rice and lossy wavelet-based compression tools (used in PICARD mission by CNES 2009), discrete cosine transform (DCT) + scalar quantization + Huffman coding (Clementine, NASA 1994), and ICER (a low-complexity wavelet-based progressive compression algorithm used in Mars mission, NASA 2003) to (12-bit) JPEG-baseline (Trace NASA1998, Solar-B JAXA2006) [1]. The key characteristics of these algorithms are relatively low encoding power consumption, coding efciency, and error resilience features. Note that all current missions, including STEREO, use 2-D monoview image compression trading off computational cost and compression perforOnboard Low-Complexity Compression of mance. Since STEREO images are essentially multiview images, with Solar Stereo Images high interview correlation, current compression tools do not provide an optimum approach. In this paper, we propose a distributed multiview Shuang Wang, Lijuan Cui, Samuel Cheng, Lina Stankovic, and image compression (DMIC) scheme for such emerging remote sensing Vladimir Stankovic setups. When an encoder can access images from multiple views, a joint coding scheme [2] achieves higher compression performance than AbstractWe propose an adaptive distributed compression solution schemes with separate coding. However, due to the limited computing using particle ltering that tracks correlation, as well as performing and communication power of space imaging systems, it is not feasible disparity estimation, at the decoder side. The proposed algorithm is to perform high-complexity power-hungry onboard joint encoding of tested on the stereo solar images captured by the twin satellites system of NASAs Solar TErrestrial RElations Observatory (STEREO) project. captured solar images. Although, intuitively, this restriction of separate Our experimental results show improved compression performance w.r.t. encoding seems to compromise the compression performance, disto a benchmark compression scheme, accurate correlation estimation by http://ieeexploreprojects.blogspot.com theory [3], [4] proves that independent tributed source coding (DSC) our proposed particle-based belief propagation algorithm, and signicant encoding can be designed as efciently as joint encoding as long as peak signal-to-noise ratio improvement over traditional separate bit-plane joint decoding is allowed. decoding without dynamic correlation and disparity estimation. The proposed DMIC image codec is characterized by low-comIndex TermsDistributed source coding, image compression, multiview plexity image encoding and relatively more complex decoding meant to imaging, remote sensing. be performed on the ground. The proposed scheme extends our previous work [5], where a joint bit-plane decoder is described, which integrates particle ltering with standard belief propagation (BP) decoding to I. INTRODUCTION perform inference on a single joint 2-D factor graph. In [5], the proOnboard data processing has been a challenging task in remote posed decoding method is used in the context of monoview coding of sensing applications due to severe computational limitations of natural video based on DCT-based distributed video coding (DVC) [6]. onboard equipment. This is especially the case in deep-space ap- In this paper, we extend the scheme to multiview image compression plications where mission spacecraft are collecting a vast amount of to further reduce the complexity we work in the pixel domain. The key images. In such emerging applications, efcient low-complexity image contributions of this paper can be summarized as follows. compression is a must. While conventional solutions such as JPEG An adaptive distributed multiview image decoding scheme, which have been used in many prior missions, the demand for increasing can estimate the blockwise correlation and disparity change beimage volume and resolution, as well as increased space resolution tween two correlated images. A BP decoder with integrated particle ltering to estimate blockwise correlation changes in the pixel domain. This extends our Manuscript received July 17, 2011; revised November 28, 2011; accepted January 17, 2012. Date of publication February 13, 2012; date of current verprevious work [5], [7] from 1-D correlation estimation to 2-D and sion May 11, 2012. This work was supported in part by the National Science from time-varying correlation estimation to spatially varying corFoundation under Grant CCF 1117886. This paper was presented in part at IEEE relation estimation. International Conference on Image Processing (ICIP-2011), Brussels, Belgium, A joint bit-plane decoder (as compared with the traditional sepSeptember 2011. The associate editor coordinating the review of this manuscript and approving it for publication was Prof. Brian D. Rigling. arate bit-plane decoder [8]), which allows the estimation of the S. Wang, L. Cui, and S. Cheng are with School of Electrical and Computer correlation and the disparity between two pixels directly rather Engineering, The University of Oklahoma at Tulsa, Tulsa, OK 74135-2512 USA than just the correlation between a corresponding pair of bits of (e-mail: shuangwang@ou.edu; lj.cui@ou.edu; samuel.cheng@ou.edu). the pixels as in [5]. L. Stankovic and V. Stankovic are with Department of Electronic and ElecWe test our lossy DMIC setup with grayscale stereo solar images obtrical Engineering, University of Strathclyde, Glasgow G1 1XW, U.K. (e-mail: lina.stankovic@eee.strath.ac.uk; vladimir.stankovic@eee.strath.ac.uk). tained from NASAs STEREO mission to demonstrate high compres[13] D. Schonberg, S. C. Draper, C. Yeo, and K. Ramchandran, Toward compression of encrypted images and video sequences, IEEE Trans. Inf. Forensics Security, vol. 3, no. 4, pp. 749762, Dec. 2008. [14] A. Kumar and A. Makur, Lossy compression of encrypted image by compressing sensing technique, in Proc. IEEE TENCON, 2009, pp. 16. [15] X. Zhang, Lossy compression and iterative reconstruction for encrypted image, IEEE Trans. Inf. Forensics Security, vol. 6, no. 1, pp. 5358, Mar. 2011. [16] A. Bilgin, P. J. Sementilli, F. Sheng, and M. W. Marcellin, Scalable image coding using reversible integer wavelet transforms, IEEE Trans. Image Process., vol. 9, no. 11, pp. 19721977, Nov. 2000. [17] D. Taubman, High performance scalable image compression with EBCOT, IEEE Trans. Image Process., vol. 9, no. 7, pp. 11581170, Jul. 2000. Color versions of one or more of the gures in this paper are available online at http://ieeexplore.ieee.org. Digital Object Identier 10.1109/TIP.2012.2187669
1[Online.] Available: sion/index.html

http://www.nasa.gov/mission_pages/stereo/mis-

1057-7149/$31.00 2012 IEEE