Multimed Tools Appl DOI 10.1007/s11042-016-3332-3

Collective blind image watermarking in DWT-DCT domain with adaptive embedding strength governed by quality metrics Hwai-Tsu Hu 1 & Ling-Yuan Hsu 2

Received: 28 July 2015 / Revised: 3 January 2016 / Accepted: 2 February 2016 # Springer Science+Business Media New York 2016

Abstract In this paper the merits of block-based blind watermarking in a composite DWTDCT domain are explored. To improve the performance in robustness and imperceptibility, the quantization index modulation (QIM) applied to DWT-DCT coefficients has been formulated in an adaptive manner, where controlling parameters are designed to minimize the bit error rates of extracted watermarks subject to a quality criterion. The targeted coefficients are chosen based upon signal analysis in combination with additional consideration of human visual characteristics. To further enhance watermarking efficiency, two collective strategies are proposed. One takes advantage of multi-bit embedding and the other modifies the norm of a vector constituted by selected coefficients. Experimental results show that, in comparison with other DWT- and/or DCT-related watermarking methods, the proposed collective schemes achieve satisfactory improvements in robustness while the imperceptibility is properly maintained. Keywords Block-based image watermarking . Discrete wavelet transform . Discrete cosine transform . Progressive quantization index modulation . Collective embedding strategy

1 Introduction Due to the advance of information technology, the access, modification and redistribution of digital multimedia data are now much easier than before. While the protection of multimedia

* Hwai-Tsu Hu [email protected]

1

Department of Electronic Engineering, National I-Lan University, No. 1, Sec. 1, Shen-Lung Rd., Yi-Lan City, Yi-Lan, Taiwan 26041, Republic of China

2

Department of Information Management, St. Mary’s Junior College of Medicine, Nursing and Management, 100, Lane 265, Sansing Rd., Sec. 2, Sansing Township, Yi-Lan, Taiwan 26644, Republic of China

Multimed Tools Appl

content increasingly became an important issue, digital watermarking was proposed as a promising resolution to prevent intellectual property infringement. Watermarking techniques are intended to insert information into digital data without introducing perceptual changes and later extract the inserted information for a variety of purposes such as content authentication, integrity verification and secret communication. There are four important concerns for digital watermarking, namely, imperceptibility, robustness, capacity and security [40]. Researchers have been trying numerous methods to improve these four aspects. Because original host images are not always accessible at the receiving side, the watermark detection is often proceeded without referring to the primordial data. Under such a blind condition, many watermarking methods suffer various degrees of deficiencies in robustness, transparency, and payload capacity. Therefore, an effective blind watermarking scheme is still urged nowadays. For the type of digital images, the watermarking process can be performed in spatial or transform domains. The spatial-domain methods are often simple but accompanied with inferior robustness. In contrast, methods emerging from transform domains such as discrete wavelet transform (DWT) [1, 2, 9, 21, 22, 36, 38, 39], discrete cosine transform (DCT) [1, 2, 5, 14, 22, 24, 25, 29, 37], discrete Fourier transform (DFT) [31, 33, 34], singular value decomposition (SVD) [4, 7, 8, 12, 26, 32] and other decompositions [27, 28] can exploit signal characteristics and human perception properties to attain better robustness and imperceptibility. The DWT and DCT are the two most popular transform domains for implementing image watermarking. The DWT is favored because it simultaneously explores spatial and spectral properties of the image [36]. By contrast, the DCT holds the advantage of excellent energy compaction for highly correlated image data. Imposing invisibility constraints is comparatively easy when working in the DCT domain [20]. It was indicated by Dowling et al. [16] in their comparison between the DWT and DCT block-based watermarking that the DCT approach is more effective at smaller block sizes while the DWT approach is probably superior if the watermarking is applied to the entire image. In [1], Al-Haj proposed using a combination of DWT and DCT to implement blind image watermarking. After applying DWT to decompose the host image into multi-resolution subbands, he divided the selected subbands into blocks of size 4 × 4. Each block was then converted to DCT representation. Binary embedding was carried out by adding uncorrelated pseudorandom sequences to the middle-frequency DCT coefficients. The watermark was extracted by calculating the correlation between the selected coefficients and pseudorandom sequences. Performance evaluation results showed that combining the two transforms could improve the watermarking algorithms that were solely based on the DWT. Several other DWTDCT based watermarking algorithms [2, 15, 18, 19, 23] were developed later based on AlHaj’s framework. The differences amongst these DWT-DCT based schemes primarily lie in the DWT decomposition levels, the subbands selected for watermarking, and the choice of DCT components in each block. Basically, they all suffered from a similar drawback that the correlation measure cannot guarantee 100 % accuracy of recovered watermarks even in the absence of attacks. To really take advantage of both DWT and DCT, we consider modulating the coefficients acquired from the DWT-DCT domain in a collective manner. More specifically, we use onelevel DWT to decompose the host image into four subbands, including one approximation subband LL and three detail subbands, namely LH, HL and HH respectively in the horizontal, vertical, and diagonal directions. The selected subbands are then partitioned into small image blocks, each of which is further transformed to DCT coefficients. After that, a novel

Multimed Tools Appl

progressive QIM (PQIM) scheme is introduced along with two collective strategies to maximize the watermarking efficiency. The main purpose of this paper is to develop a block-based watermarking technique applicable in the DWT-DCT domain. In principle, a block-based scheme deals with every block separately, thus providing several advantages. For example, we can exploit the distinct features (such as local luminance and texture) of each individual image block to improve the robustness of the embedded watermark and/or enhance the visual quality of the watermarked image. We can use a different secret key in each block to reinforce the security of the watermark. The amount of information bits may also vary from one block to another. Moreover, the block-based schemes are very suitable for applications dedicated to the regions of interest (ROI). Either the blocks inside or those outside the ROIs can be selected for watermarking. The rest of this paper is organized as follows. Section 2 gives a brief discussion about the block-based transform-domain schemes for binary image watermarking. Section 3 describes the feature parameters adopted for image watermarking. Section 4 presents the PQIM technique. Section 5 provides two collective strategies to enhance the watermark robustness. Section 6 explains our approach for seeking ideal controlling parameters. Experimental results and analysis are given in Section 7. Finally, Section 8 draws our conclusions.

2 Previous DWT- and DCT-based methods for binary image watermarking In image watermarking, the inserted watermark can be either a pseudorandom sequence or binary bits. For the type of pseudorandom sequences, almost all of the watermarking methods are developed directly or indirectly from Cox et al.’s spread spectrum technique [13]. For the type of binary bits, there are generally two ways to implement the watermarking. One is to map the chosen coefficient to a dichotomized field according to the bit value. The quantization index modulation (QIM) proposed by Chen and Wornell [10, 11] is the most common technique. The other way is to manipulate the coefficients in paired groups so that each binary bit is manifested by the relationship of the paired groups [1, 2, 6, 22, 32, 37]. Our focus in this study is particularly placed on binary image watermarking due to its practical applications. Various kinds of binary watermarking approaches had been attempted with success in recent years. In [39], Zhang et al. directly embedded the watermark in the DWT domain. They analyzed the characteristics of the DWT detail coefficients of the host image and adaptively modified involved coefficients based on the mean and variance of the detail subbands. In [9], Chen et al. embedded binary bits into the second level LH subband using an optimized quantization scheme, which aimed at connecting the performance and amplitude modulation. Unlike the traditional way of single-coefficient quantization, they particularly modulated the weighted sum of the magnitudes of consecutive coefficients drawn from the selected subband. The useful properties of the DWT can also be exploited via SVD. In a scheme proposed by Bao and Ma [4], the LL subband of one-level DWT was first divided into non-overlapping blocks of size 4 × 4. The SVD was then brought in to factorize each block into three component matrixes, namely a left singular matrix U, a diagonal singular matrix D, and a right singular matrix V. Next, they applied standard QIM to modulate the vector subsuming all singular values. The quantization step was designed to adapt to the mean and standard deviation obtained from each block. In [7], Chang et al. manipulated the relationship between

Multimed Tools Appl

two eigenvalues to achieve binary embedding. Chang et al. [8] demonstrated another feasibility by altering the magnitudes of two elements in the first column of U to satisfy a predefined inequality relationship. This approach was modified later by Fan et al. [17] and Su et al. [32] such that the visible distortion can be compensated through the adjustment of V. For the category of DCT-based methods, the host image is commonly divided into nonoverlapping blocks of size 8 × 8, each of which undergoes DCT conversion individually. The watermark embedding is then carried out by altering DCT coefficients according to certain rules and/or required conditions. In [25], Lin et al. embedded binary bits into selected blocks using standard QIM. Patra et al. [29] employed the Chinese remainder theorem (CRT) technique to modulate a randomly picked coefficient in each DCT block. Wang and Pearmain [37] used relative modulation to achieve binary embedding, apart from the exploitation of inter-block correlation. Inside each DCT block, they estimated specific DCT coefficients by referring to the DC values at surrounding blocks. The watermark embedding was done by shifting the designated DCT coefficient to a position above the estimated value if the watermark bit wb was B1^ or to a position below otherwise. Another way of exploiting the inter-block correlation was proposed by Das et al. [14], who developed a watermarking algorithm using the difference between two designated DCT coefficients in adjacent blocks. The targeted DCT coefficient was modified to bring the difference to a specified range. The idea of jointly using the DWT and DCT has also been examined in the past. In [1], AlHaj applied DWT to the HL subband matrix after taking the first DWT of the host image. The second DWT applied to the HL matrix results in four subbands expressed in form of HL1 ↦ LL2, HL1 ↦ HL2, HL1 ↦ LH2 and HL1 ↦ HH2, where the subscript index denotes the decomposition level. The subband matrix HL1 ↦ HL2 was then partitioned into non-overlapped block of size 4 × 4, each of which underwent two-dimensional (2-D) DCT to obtain a transformed matrix. The actual binary embedding was implemented using an additive scheme. Given that X denotes the vector whose elements are drawn from the mid-band DCT coefficients, the watermark embedding in [1] was attained via
0 X þ α Ψ 0 ; if wb ¼ 0; ð1Þ X ¼ X þ α Ψ 1 ; if wb ¼ 1; where Ψ0 and Ψ1 are two uncorrelated pseudorandom sequences. The variable α controls the embedding strength. In [2], Amirgholipour and Naghsh-Nilchi adopted a similar DWT-DCT framework except that they utilized three-level DWT. The subband matrices selected for embedding a watermark of size 32 × 32 included HL1 ↦ HL2 ↦ HL3, HL1 ↦ LH2 ↦ HL3, LH1 ↦ HL2 ↦ HL3 and LH1 ↦ LH2 ↦ HL3. Laskar et al. [23] chose exactly the same subbands as in [2] to perform digital watermarking for ownership verification of digital images. In [19] and [15], both the algorithms embedded the watermark in HL1. For the one presented in [19] the pseudorandom sequence was added to the mid-band DCT coefficients, while the addition was conducted at the low-frequency range for the one presented in [15]. Because most of the image energy is concentrated at lower frequencies, Feng et al. [18] considered embedding the watermark into LL1 to increase robustness. Nevertheless, the watermark embedding was still performed within the middle-frequency range of each DWT-DCT block. Overall, the main differences for the algorithms presented in [1, 2, 15, 18, 19, 23] consist in DWT decomposition levels and selected subbands for watermarking. Recently, another DWT-DCT watermarking technique was tried by Kalra et al. [22]. They first located a middle-frequency coefficient in each block and then adaptively assigned a

Multimed Tools Appl

positive value if wb = 1 and a negative one if wb = 0. The embedding strength, as signified by the assigned value, is determined by taking consideration of the frequency location of the coefficient and the variance of the host image.

3 Data analysis for the extracted DWT-DCT coefficients In this study, we use one-level Haar wavelet transform [30] to decompose a host image into LL, HL, LH and HH subbands. Because the allowable embedding strength of a watermark is normally proportional to signal energy and most of the energy of a nature image is concentrated at the low-frequency region, we therefore choose the LL subband to embed the watermark. However, human eyes are also sensitive to the changes in the low-frequency subband. Embedding watermarks in the LL subband may degrade image quality significantly. A compromise can be settled by embedding the watermark in the middle-frequency range of the LL subband. More specifically, we divide the LL subband matrix into non-overlapping blocks of size 4 × 4 and then convert each block separately to the spectral domain using 2-D DCT. Each of the resulting DWT-DCT blocks thus consists of one DC component on the upper left corner and fifteen AC components distributing over two dimensions. Figure 1 depicts the array arrangement of a DWT-DCT block. The watermark embedding can be done by manipulating the designated coefficients in each block. We note here that every DWT-DCT block covers an area of 8 × 8 pixels in the host image. It is the exact size adopted in the JPEG compression standard and commonly used for image watermarking as well. The DWT-DCT transform offers two advantages. First, as the watermarking process merely involves the LL subband, the remaining subbands can be used for other purposes such as developing a content-based algorithm to resist desynchronization attacks. Second, since the required computation of 2-D DCT for a block of size m × n is on the order of mn(log m + log n) [35], reducing the block size down to 4 × 4 can accelerate the processing speed to a large degree. To grasp the fundamental characteristics of DWT-DCT blocks, we examine twenty-four gray images of size 512 × 512 in advance. These images can be classified into 6 categories, including “humans”, “animals”, “flowers”, “fruits and vegetables”, “scenes” and “transportation vehicles”. Figure 2 shows these images in a 6-by-4 grid. The statistical analysis with respect to the DWT-DCT coefficients acquired from test images confirms that most signal

DC(0,0)

AC(0,1)

AC(0,2)

AC(0,3)

AC(1,0)

AC(1,1)

AC(1,2)

AC(1,3)

AC(2,0)

AC(2,1)

AC(2,2)

AC(2,3)

AC(3,0)

AC(3,1)

AC(3,2)

AC(3,3)

Fig. 1 Arrangement of a 4x4 DWT-DCT block

Multimed Tools Appl

Fig. 2 Gray images for test

information is concentrated in the low-frequency region. When the images are compressed using the JPEG standard, the images suffer more distortion if the quality factor (QF) declines. In particular, the variations are most pronounced at high frequencies but roughly remain at a similar level at middle frequencies. As far as the robustness is concerned, it is preferable to have the watermark embedded in the DWT-DCT coefficients with large magnitudes. One obvious target is the DC component since its magnitude is significantly larger than others’. Unfortunately, human eyes are very sensitive to the change of DC components. Hence we turn our attention to other AC components. Because the low-frequency components are less perturbed in comparison with

Multimed Tools Appl

the high-frequency ones, we therefore conclude that the watermark bits had better be inserted into the coefficients in the plus-second and plus-first anti-diagonals of the DWT-DCT blocks.

4 Adaptive quantization index modulation In light of the analysis in the previous section, we choose AC(1, 0), AC(0, 1) and AC(1, 1) in each DWT-DCT block as the feature parameters for watermark embedding. As mentioned in the introduction, the QIM [10, 11] has been considered an efficient means to embed binary information. The basic idea of QIM is to quantize the coefficient amplitude to an odd index if the watermark bit is B1^ and an even index if the watermark bit is B0^. The only problem left here is how to set the quantization step size. Note that the embedded watermark may be either perceptible by human eyes if the step size is too large or vulnerable to attacks if the size is too small. A favorable choice is to adapt the quantization step size based on signal characteristics. Based upon our experimental observations, the blocks in smooth areas often have small AC components, whereas the ones in highly textured areas are associated with large AC components. Since the human visual system is rather sensitive to the noise in smooth areas, we decrease the quantization step size when the AC components are small and gradually increase the step size when the AC components become large. Eventually, we quantize the DWT-DCT coefficients in a progressive manner such that the quantization step expends more as the quantized index increases. Let the nth quantization step Δn be of the form: Δn ¼ Δ þ nδ;

ð2Þ

where Δ is a fixed value and δ is the increment between adjacent steps. By assuming that ρ is one of the DWT-DCT coefficients among AC(1, 0), AC(0, 1) and AC(1, 1), the magnitude of ρ will be equivalent to S step units via an integral equation: Z S ðΔ þ xδÞdx jρj ¼ 0 ð3Þ S2 ¼ ΔS þ δ : 2 Solving the above quadratic equation gives qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi −Δ þ Δ2 þ 2δjρj S¼ : ð4Þ δ To embed a binary bit wb ∈ {0, 1}, we alter S as    8  > > 2 S þ 1 −wb ; if 2 S þ 1 ≥S; <  2  2  S^ ¼ Sþ1 Sþ1 > > :2 þ wb ; if 2 < S; 2 2

ð5Þ

^ where ⌊ · ⌋ denotes the floor function and Ŝ stands for the quantized index of the adjusted ρ such that 0

1 2 ^ S ^ ¼ sgnðρÞ⋅@ΔS^ þ δ A ρ 2

with sgn(·) symbolizing the sign function.

ð6Þ

Multimed Tools Appl

After modifying the designated DWT-DCT coefficient, we take the inverse DCT block-byblock to restore the LL subband matrix. The watermarked image can be attained by performing one-level inverse DWT with the modified LL subband and three original detail subbands taken as inputs. Figure 3 illustrates the procedural flow of the proposed watermarking scheme. Given that the size of the host image is L × K, the maximum allowable size of the watermark becomes nL × nK, where nL ≤ ⌊L/8⌋ and nK ≤ ⌊K/8⌋. To enhance the security, we employ the Arnold transform [3] to scramble the watermark before activating the embedding process. The rest embedding procedure consists of a one-level DWT and nL × nK times of DCT along with the progressive QIM for each DWT-DCT block. As for watermark detection, we apply one-level DWT to the watermarked image, partition the LL subband matrix into blocks of size 4 × 4, and subsequently take the DCT of each block. The magnitude of the recovered DWT-DCT coefficient ~ ρ is converted to an index S~ using the PQIM, i.e.

S~ ¼

Fig. 3 Watermark embedding procedure for the block-based DWT-DCT scheme

rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi     −Δ þ Δ2 þ 2δ~ ρ δ

:

ð7Þ

Host image 1-level DWT Paron LL subband into 4x4 blocks

DCT

Watermark Arnold Transform

Embed bit(s) using PQIM

IDCT For mul-bit embedding End of Blocks

IDWT

Watermarked image

Multimed Tools Appl

The watermark bit is classified as B1^ if S~ is closer to an odd integer and classified as B0^ otherwise: j k ~ b ¼ S~ þ 0:5 mod 2; ð8Þ w where Bmod^ denotes the modulus operation. Once all the watermark bits are obtained and arranged in a matrix of size nL × nK, an inverse Arnold transform is employed to reestablish the watermark.

5 Collective watermarking One possible way to exert the advantage of the DWT-DCT framework is the use of multiple coefficients collectively. There are two strategies considered in this study. The first one embeds each watermark bit repeatedly into multiple places and determines the final outcome from all collected bits using a majority vote. The second one simply takes the selected coefficients as a whole and varies the involved coefficients jointly according to the embedding rule. To implement the first strategy, we scramble the binary watermark into three unrecognizable versions using the Arnold transform. These three scrambled watermarks are then embedded in three locations, namely AC(0, 1), AC(1, 0) and AC(1, 1), of the DWT-DCT blocks. The procedure for multi-bit embedding and extraction is identical to that adopted in the single-bit case except that the retrieved bits must receive an extra judging process. More specifically, once three candidate bits are restored by the inverse Arnold transform, the final verdict for the bit value is the one supported by more than half the candidate bits. The chance of obtaining an erroneous result depends on the error probabilities of the candidates. In our case, the error probability of the final result, termed perror, can be derived from perror ¼ p1 p2 p3 þ ð1−p1 Þp2 p3 þ p1 ð1−p2 Þp3 þ p1 p2 ð1−p3 Þ;

ð9Þ

where p1, p2 and p3 denote the error probability of each individual candidate. In general, perror will be smaller than any of p1, p2 and p3 if all three are less than 0.5. Our second collective strategy regards the selected coefficients as a vector. That is, the watermark embedding is attained by modulating the vector norm η according to the PQIM. qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ð10Þ η ¼ AC 2 ð1; 0Þ þ AC 2 ð1; 0Þ þ AC 2 ð1; 1Þ: To avoid the condition of dividing zero in subsequent computation, we replace these three AC components by small random numbers if all of them are zero. Once the modulation is done using the PQIM presented in Section 3, the AC components are scaled by  c ði; jÞ AC

ði; jÞ∈fð0;1Þ;ð1;0Þ;ð1;1Þg

¼

^η ACði; jÞ; η

ð11Þ

where ^ η is the modulated value of η. During watermark extraction, we derive the vector norm of the gathered AC components for every DWT-DCT block and determine the watermark bit by examining the quantized index of the vector norm. Note that, under the PQIM, a large ^η is theoretically more capable of withstanding attacks. By bundling more AC components

Multimed Tools Appl

together as a vector, we offer the embedded bit a better chance to survive the attacks as long as the vector norm holds a sufficiently large magnitude.

6 Embedding strength control As mentioned earlier, the PQIM involves two parameters, namely, Δ and δ. In general, increasing Δ and δ will contribute to the robustness of the watermark but harm the perceived quality. The first important step is therefore to choose suitable Δ and δ for the DWT-DCT coefficients in each specific position. To attain a trade-off between robustness and imperceptibility, we have conducted a pilot test with the six gray images in the first column of Fig. 2 taken as tentative objects. The watermark in this study was a 64 × 64 binary image logo, of which the numbers of B1 s^ and B0 s^ were deliberately arranged to be equal. The most appropriate combination of Δ′ and δ′ is chosen as the one that minimizes the overall error rate summed up from nP(=6) testing objects in the presence of nA typical attacks under a prespecified quality constraint, i.e.

0

Δ 0 δ

¼ min arg

nP X nA   X ^ BERði; jÞ W; W

ð12Þ

Δ∈GΔ i¼1 j¼1 δ∈Gδ

subject to  Qcomp

  9 8 = < PSNR I; ^I    ; 1  MSSIM I; ^I ≥qthd ; I; ^I ¼ max ; : 40

ð13Þ

where GΔ and Gδ respectively represent two numeric sets used in the search of suitable Δ and δ. Qcomp(I, Î) denotes a composite quality metric measured between the original image I and watermarked image Î. The three abbreviations, namely BER, PSNR and MSSIM, stand for the Bit Error Rate, Peak Signal-to-Noise Ratio and Mean Structural SIMilarity, respectively. The BER between the original watermark W = {Wij} and extracted watermark Ŵ = {Ŵij} is defined as nL X nK X



 ^ ¼ BER W; W

^ ij W i j ⊕W

i¼1 j¼1

nL  nK

;

ð14Þ

where ⊕ denotes the exclusive-OR operation. The PSNR assesses the discrepancy between I = {Iij} and Î = {Îij} as compared to the maximum possible power:   PSNR I; ^I ¼ 10log10

2552 L K 1 X X  ^ 2 I i j −I i j L  K i¼1 j¼1

½dB;

ð15Þ

Multimed Tools Appl

The MSSIM denotes the mean value of evaluated quality based on the SSIM metric. N   1X MSSIM I; ^I ¼ SSIM ðxi ; yi Þ; N i¼1

ð16Þ

where xi is the ith local region of the original image I, yi represents its counterpart in the distorted image Î, and N denotes the total quantity of local regions. The SSIM metric given underneath is applied to I and Î locally within an 8 × 8 window. 

  2μxi μyi þ ðk 1 RÞ2 2σxi yi þ ðk 2 RÞ2  ; SSIM ðxi ; yi Þ ¼  μ2xi þ μ2yi þ ðk 1 RÞ2 σ2xi þ σ2yi þ ðk 2 RÞ2

ð17Þ

where μzi and σ2zi respectively denote the mean and variance of zi ∈ {xi, yi}. σxi yi is the covariance between xi and yi. R signifies the dynamic range of pixel values and it is equal to 255 for grayscale images. k1 and k2 are two small values introduced to ensure the stability of the SSIM metric. They are set up as 0.01 and 0.02 by default. In the formulation of Eq. (17), the mean and variance can be viewed as the estimates of luminance and contrast, while the covariance measures the structural tendency between two local regions. The resulting Qcomp(I, Î)’s are designedly distributed between 0 and 1 with 1 indicating a perfect match. Based on the formulation in Eq. (13), watermarked images with PSNRs below 40 dB will suffer more penalties. Eq. (12) thus aims at seeking the best combination of (Δ′, δ′) that achieves the minimum BER while still maintaining the image quality beyond a specified level qthd. The search for Δ′ and δ′ was experimentally performed over a mesh grid constituted by two finite arithmetic sequences, i.e., GΔ and Gδ. For single DWT-DCT coefficient modulation in AC(0, 1), AC(1, 0) and AC(1, 1), the search ranges were given as
 8 1 1 1 > > < GΔ ¼ 14; 14 ; 15; 15 ; ⋯; 20 ; 21 ; 2 2 2
 1 3 1 1 3 > > : Gδ ¼ 3 ; 3 ; 4; 4 ; 4 ; ⋯; 6 ; 7 : 2 4 4 2 4

ð18Þ

The quality threshold, termed qthd in the inequality expression (13), is chosen as 0.9875. For the vector norm η in the second strategy, the two search ranges were changed to
 8 1 1 1 > > < GΔ ¼ 27; 27 ; 28; 28 ; ⋯; 35 ; 36 ; 2 2 2
 1 1 3 1 1 > > : Gδ ¼ 5; 5 ; 5 ; 5 ; 6; ⋯; 9 ; 9 : 4 2 4 4 2

ð19Þ

The threshold qthd is accordingly lowered to 0.92, which is a level similar to the effect by altering the coefficients in AC(0, 1), AC(1, 0) and AC(1, 1) concurrently. In the pilot test we only considered three representative attacks, which were the JPEG compression with quality factor = 20, JPEG 2000 with compression ratio = 8, and median filter with a 3 × 3 mask. After settling all conditions, we executed the search procedure, i.e., Eq. (12), straightforwardly. Table 1 lists the resultant Δ′ and δ′ for the selected DWT-DCT coefficients.

Multimed Tools Appl Table 1 Parameters adopted in the PQIM Δ

δ

AC(0, 1)

18.50

4.75

AC(1, 0)

18.00

4.75

AC(1, 1) η = |(AC(0, 1), AC(1, 0), AC(1, 1))|

16.50 31.50

6.25 6.25

Parameter Position

7 Experimental results and performance comparison The performance of the proposed scheme was compared with that obtained from fourteen other block-based methods mentioned in Section 2. The testing materials included all the 24 images shown in Fig. 2. To make the performance evaluation on an equitable basis, we unified the payload capacity as 1/64 bit per pixels. That is, every 8 × 8 block contains one bit after the embedding process. Note that the watermark size considered in [1] was 32 × 32, which was only one-fourth of the one adopted in this study. Hence our implementation took into account another three second-level subband matrices, namely HL1 ↦ HH2, HH1 ↦ HL2 and HH1 ↦ HH2. Likewise, the implementation of the method in [2] additionally subsumed HL 1 ↦ HL 2 ↦ XX 3 , HL 1 ↦ HL 2 ↦ XX 3 , LH 1 ↦ HL 2 ↦ XX 3 and LH 1 ↦ LH 2 ↦ XX 3 with XX ∈ {LL, HL, LH, HH} so as to accommodate a watermark of size 64 × 64. Moreover, the embedding strength α was fixed as 8.5 to render a comparable signal-to-noise ratio specified in [2]. Similarly, for the algorithms considered in [18, 19] and [22], we have deliberately quadrupled the capacity and adjusted the embedding strength factor to retain proper imperceptibility. To assess the image quality distortion due to various watermarking methods, we provide the statistics of measured PSNRs and MSSIMs in Table 2. The average PSNRs obtained from the proposed scheme with only one single AC component involved are all above 42 dB. The corresponding MSSIMs are around 0.986, which implies very good image quality. The CRT-based method [29] can also maintain a comparable quality level. In contrast, the average PSNRs achieved by the methods in [39] and [14] are noticeably inferior. However, their MSSIMs are not necessarily among the worst. Though the average PSNRs for the methods in [4] and [25] are both around 41 dB, the obtained MSSIMs indicate that the perceivable distortion for the DWT-SVD-based method in [4] appears much less than that obtained from the DCT-based method in [25]. The high MSSIMs associated this DWT-SVD method can be attributed to the preservation of the internal structure in each block, since the singular values are adjusted in an all-round manner. Finally, somewhat to our surprise, the most unfavorable MSSIMs accompany with the DWT-DCT-based methods in [1, 2, 18, 19]. It is reminded here that, no matter which watermarking method is adopted, the resultant image suffers various degrees of quality deterioration. The image is subjected to more distortion if we simultaneously impose the embedding process to three DWTDCT coefficients. The average PSNRs of the two collective watermarking schemes

0.958

0.007

Std

1.911

std

mean

36.37

0.015

0.963

0.497

39.66

DWT

DWT

mean

Ref. [9]

Ref. [39]

0.002

0.993

0.664

41.23

DWTSVD

Ref. [4]

Ref. [32]

0.005

0.983

3.178

38.62

0.004

0.983

1.781

40.81

0.009

0.971

0.110

40.85

0.004

0.989

0.297

42.22

DCT

Ref. [25] Ref. [29]

DWT- SVD DWT- SVD DCT

Ref. [7]

0.007

0.973

2.068

37.30

DCT

Ref. [37]

0.021

0.932

0.000

37.60

DCT

Ref. [14]

Ref. [2]

Ref. [18]

Ref. [19]

Ref. [22]

Multi-bit

The proposed scheme

0.021

0.932

0.000

37.60

0.020

0.936

0.000

37.60

0.021

0.932

0.000

37.60

0.021

0.932

0.000

37.60

0.032

0.942

2.904

38.48

0.011

0.960

0.518

37.79

Vector

<0.004, 0.004, 0.004 >

<0.986, 0.986, 0.986 >

<0.678, 0.483, 0.540 >

0.009

0.965

0.518

<42.19, 42.29, 38.18 43.34 >

DWT- DCT DWT- DCT DWT- DCT DWT- DCT DWT- DCT DWT-DCT

Ref. [1]

Note: is interpreted as a set within which each element represents the result when only one AC coefficient is utilized for watermarking in each DWTDCT block

MSSIM

PSNR

Measure Method Domain

Table 2 Measured PSNRs and MSSIMs in terms of mean and standard deviation (std)

Multimed Tools Appl

Multimed Tools Appl

turns out to be 37.79 and 38.18 dB for the multi-bit embedding and vector norm modulation respectively, indicating that the alterations due to the watermarking process are still at an acceptable level. The corresponding MSSIMs are 0.960 and 0.965 in average for these two collective schemes, suggesting that the perceivable distortion caused by applying the PQIM to the whole group is slightly smaller than that by separate modulation. As for the DWT-DCT methods developed in [1, 2, 18, 19], the resulting MSSIMs are relatively low even though the PSNRs are purposely controlled around 37.60 dB. The cause of such inferior results is conceivably ascribed to the fact that the alteration in the DWT-DCT block is apt to perturb the structural dependency among spatially close pixels. Embedding a watermark four times its originally designed capacity further deteriorates the MSSIM. The same explanation suits the conditions observed in another DWT-DCT method [22], of which the average PSNR is 38.48 dB and the resulting MSSIM is merely 0.936. The robustness of the proposed watermarking scheme under commonly encountered attacks was evaluated using the BER defined in Eq. (14). The attack types included: 1. JPEG compression: The JPEG with quality factor (QF) chosen from {80, 40, 20} is applied to the test image. 2. JPEG2000 compression: The JPEG2000 with compression ratio (CR) chosen from {2, 4, 8} is applied to the test image. 3. Gaussian noise (0.001): The test image is corrupted by Gaussian noise with the variance set as 0.001 of the full scale. 4. Salt-and-pepper noise (1 %): The test image is corrupted by the salt-and-pepper noises with 1 % intensity. 5. Median filter (3 × 3): A median filter with a 3 × 3 mask is applied to the test image. 6. Gaussian filter (3 × 3): A Gaussian filter with a 3 × 3 mask is applied to the test image. 7. Scaling correction (25 %): The size of the test image is first shrunk from 512 × 512 to 256 × 256 pixels and subsequently enlarged from 256 × 256 to 512 × 512 pixels. 8. Rotation correction: The test image is first rotated counterclockwise by 45Å and then 45Å back to the original position. 9. Cropping (25 %): The test image is cropped 25 % on the upper-left corner. 10. Level shifting/ brightening (+20): The operation adds 20 to each pixel value of the test image. 11. Level shifting/ darkening (−20): The operation subtracts 20 from each pixel value of the test image. Table 3 presents the average BERs obtained from various watermarking methods. To give a general view of the performance of every examined method, we depict the retrieved watermarks in Table 4 for intuitive inspection and comparison. As shown in Tables 3 and 4, the two collective DWT-DCT-based schemes evidently outperform the other fourteen methods for JPEG and JPEG 2000 image compression. The methods developed in [4, 9, 14, 22, 25] and [37] also show certain resistance to compression attacks, but their performances are inferior to the proposed ones. Especially, the robustness of the methods in [22, 37] and [14] is gained at the cost of quality degradation. The method in [4] cannot survive the levelshifting attacks, i.e. types 10 & 11, which however cause no problems to the DCTbased method in [25].

24.73

13.54

36.52

5.18

29.81

33.84 11.99

1.11

2.26

4

5

6

7

8 9

10

11

0.16 5.06

2/ CR 2 2/ CR 4

12.09

38.23

1/ QF 20

3

21.62

1/ QF 40

2/ CR 8

0.00

2.78

1/ QF 80

0.00

2.82

1.17

34.81 12.43

27.35

4.17

27.27

26.42

15.69

13.80

0.07 1.15

29.70

10.46

0.33

0.00

48.73

48.11

29.52 11.99

4.78

1.28

11.10

26.76

18.69

8.57

0.04 1.06

42.32

7.73

0.11

0.01

3.60

1.37

30.47 11.99

27.64

14.08

31.28

11.71

23.52

28.32

3.43 11.15

36.85

25.93

7.66

0.00

4.09

2.07

29.82 11.99

22.98

6.06

30.02

22.59

35.50

25.20

0.13 4.86

33.90

25.91

7.30

0.00

3.18

1.35

26.12 11.99

7.65

0.91

18.50

17.67

13.81

15.50

0.01 1.33

12.02

0.08

0.02

0.00

67.55

68.19

32.30 11.99

15.07

8.51

22.03

19.80

22.23

20.62

3.73 9.57

27.91

20.18

5.35

0.00

2.77

1.07

17.10 12.25

9.73

1.87

14.43

12.03

12.94

14.83

0.29 2.11

33.14

2.54

0.36

0.00

2.41

1.03

17.93 13.14

8.91

1.47

13.16

17.82

10.82

10.27

0.19 1.11

25.71

5.90

0.24

3.39

4.14

3.64

24.22 15.09

39.13

4.17

27.90

11.75

5.52

16.17

3.38 4.94

45.92

35.42

18.93

3.73

4.20

4.11

26.07 15.70

46.15

4.21

49.87

13.32

6.17

15.15

3.86 5.37

44.20

31.78

14.90

10.55

11.70

11.04

17.09 21.07

14.74

11.93

15.92

15.70

12.75

17.38

10.96 11.93

25.83

13.85

12.54

5.7

6.89

6.08

26.17 17.39

38.56

8.65

60.66

11.5

8.08

15.36

5.93 6.56

41.6

23.68

8.13

0.00

0.67

0.14

23.35 13.01

29.66

0.44

51.14

7.43

1.79

9.43

0.01 0.58

42.26

27.08

2.52

0.79

0.20

10.41 7.81

0.58

0.00

3.32

11.35

5.74

3.77

0.00 0.07

0.24

0.01

0.00

0.00

3.28

1.23

13.94 11.99

1.09

0.04

4.45

14.44

5.67

3.06

0.01 0.17

0.17

0.06

0.02

0.00

Multi-bit Vector

Ref. [39] Ref. [9] Ref. [4] Ref. [7] Ref. [32] Ref. [25] Ref. [29] Ref. [37] Ref. [14] Ref. [1] Ref. [2] Ref. [18] Ref. [19] Ref. [22] Proposed

None

Attack type

Method

Table 3 Average BERs (in percentage) of recovered watermarks for compared schemes

Multimed Tools Appl

Multimed Tools Appl Table 4 Retrieved watermarks from the watermarked BLena^ images in the presence of various attacks Method Aack type

Ref.

Ref.

Ref.

Ref.

Ref.

Ref.

Ref.

Ref.

Ref.

Ref.

Ref.

Ref.

Ref.

Ref.

Proposed

[39]

[9]

[4]

[7]

[32]

[25]

[29]

[37]

[14]

[1]

[2]

[18]

[19]

[22]

Mul-bit Vector

None 1/ QF 80 1/ QF 40 1/ QF 20 2/ CR 2 2/ CR 4 2/ CR 8 3 4 5 6 7 8 9 10 11

As also revealed by Table 3, the DWT-DCT methods developed in [1] and [2] roughly achieve the same degree of performance. Both of them demonstrate relatively good robustness against noise attacks (i.e., types 3 & 4). Nevertheless, they suffer the drawback that the retrieved watermark cannot be entirely accurate even in the absence of attacks. Selecting different DWT subbands and DCT frequency ranges, such as the cases in [18] and [19], does not seem to contribute obvious improvement in BERs. The fault can be traced back to the formula used in Eq. (1). Note that, through Linear Algebra, each X in Eq. (1) can be represented as a weighted combination of three components like X ¼ aΨ 0 þ bΨ 1 þ cΨ orth ;

ð20Þ

where Ψorth is presumably orthogonal to both Ψ0 and Ψ1. If the relationship a > α + b occurs in a situation where wb = 1, the retrieved watermark bit will be interpreted as 0 instead. An analogous false interpretation can be observed in the occasion when b > α + a and wb = 0. One way to alleviate such a drawback is by raising α to a significantly large value. Unfortunately, the increase of α will impair the quality of watermarked images. In our experiments, the quality impairment is further exacerbated by a high payload capacity since more subbands are subjected to modification. By contrast, the third DWT-DCT method presented in [22] utilizes the plus or minus sign of specific subband components to indicate binary bits. This kind of approach exhibits excellent robustness against Gaussian noise and salt-and-pepper noise but completely

Multimed Tools Appl

fails in the attack of a 3 × 3 median filter, suggesting that the median filter is more likely to alter the signs of DCT-DWT coefficients. As for the two proposed collective DWT-DCT schemes, their performances in the robustness tests are quite conspicuous. Except for noise corruption and level shifting, the multi-bit embedding strategy often achieves the lowest BER among the methods considered in this study. The vector norm modulation usually ranks the second best. It even excels the multi-bit approach in the cases of high-ratio image compression and Gaussian noise corruption. One particular advantage of the multi-bit embedding is that it can partially remedy the information loss due to cropping. Theoretically, in a situation where all blocks are fully embedded with information bits, cropping one fourth of the watermarked image results in a loss of 25 % information. The resulting BERs are expected to be very close to 12.5 %, since around a half out of the 25 % are luckily guessed. With the employment of the multi-bit embedding strategy, the watermark bits are scattered to different places of the image via the Arnold transform. It is therefore possible to retrieve these bits as long as they are not entirely obliterated. In fact, the proposed collective schemes share certain similarities with the four DWT-DCT based methods in [1, 2, 18, 19]. Basically, they all employ the DWT to perform multi-resolution spectral and spatial analysis, while the DCT is used to extract the frequency features of the partitioned blocks in designated subbands. However, the proposed schemes differ from the other four in several aspects. First, the watermarks are embedded in different subband areas. Second, during the watermark embedding the proposed ones involve three DCT coefficients with relatively large and steady magnitudes, whereas the four in [1, 2, 18, 19] comprise the midband coefficients without concern for their magnitudes. Third, the proposed schemes adopt the progressive QIM to embed binary bits, while the other four use additive watermarking. It is these differences that make the distinction of the resulting performance. Note also that Bao and Ma’s method [4] required preserving adaptive quantization steps for all blocks. In their method, the steps were encoded using vector quantization and regarded as secrete keys. This kind of approach is not a blind watermarking method in a strict sense as it demands the preservation of additional information derived from the original image. By contrast, all the other methods conform to the basic requirement of blind watermarking.

8 Concluding remarks A PQIM technique has been proposed and incorporated into a block-based watermarking scheme performed in DWT-DCT domain. By analyzing the statistic properties of DWT-DCT blocks, we select three AC components, namely AC(0, 1), AC(1, 0) and AC(1, 1), for performing watermark embedding based on the PQIM. To maintain the watermarked image quality at an adequate level, the controlling parameters for the PQIM are identified through searching possible ranges subject to a quality criterion that takes both PSNR and MSSIM into account. Two collective

Multimed Tools Appl

strategies are proposed to enhance the robustness. The first one replicates the watermark and embeds multiple bits in each block. Once the multiple embedded bits are recovered, the final bit value is determined by a majority vote. The second strategy regards AC(0, 1), AC(1, 0) and AC(1, 1) as a vector and modulates the vector norm according to the PQIM. The experimental results confirm that these two collective DWT-DCT schemes outperform the other fourteen block-based methods developed in the DWT, SVD and/or DCT domains, while the resulting image quality still remains at an adequate level. Despite its superiority in robustness and imperceptibility, the proposed schemes cannot withstand desynchronization attacks at the present stage. However, several available approaches may be employed to remedy this deficiency [41]. Since our schemes are implemented in the LL subband of the first level DWT, such an arrangement allows us to make flexible use of other subbands to cope with desynchronization issues without concerning the inference from the proposed schemes. This part of study will be our future research plan. Acknowledgments This work was supported by the Ministry of Science and Technology, Taiwan, ROC, under Grant MOST 103-2221-E-197-020.

References 1. Al-Haj A (2007) Combined DWT-DCT digital image watermarking. J Comput Sci 3(9):740–746 2. Amirgholipour SK, Naghsh-Nilchi AR (2009) Robust digital image watermarking based on joint DWTDCT. Int J Digit Content Technol Applic 3(2):42–54 3. Arnold VI, Avez A (1968) Ergodic problems of classical mechanics. the mathematical physics monograph series. Benjamin, New York 4. Bao P, Ma X (2005) Image adaptive watermarking using wavelet domain singular value decomposition. IEEE Trans Circ Syst Video Technol 15(1):96–102 5. Barni M, Bartolini F, Cappellini V, Piva A (1998) A DCT-domain system for robust image watermarking. Signal Process 66(3):357–372 6. Bender W, Gruhl D, Morimoto N, Lu A (1996) Techniques for data hiding. IBM Syst J 35(3.4):313–336 7. Chang C-C, Hu Y-S, Lin C-C (2007) A digital watermarking scheme based on singular value decomposition. In: Chen B, Paterson M, Zhang G (eds) Combinatorics, algorithms, probabilistic and experimental methodologies, vol 4614. lecture notes in computer science. Springer, Berlin Heidelberg, pp 82–93 8. Chang C-C, Tsai P, Lin C-C (2005) SVD-based digital image watermarking scheme. Pattern Recogn Lett 26(10):1577–1586 9. Chen S-T, Huang H-N, Kung W-M et al. (2015) Optimization-based image watermarking with integrated quantization embedding in the wavelet-domain. Multimed Tools Appl:1–19 10. Chen B, Wornell GW (2001) Quantization index modulation methods for digital watermarking and information embedding of multimedia. J VLSI Signal Process Syst Signal Image Video Technol 27(1):7–33 11. Chen B, Wornell GW (2001) Quantization index modulation: a class of provably good methods for digital watermarking and information embedding. IEEE Trans Inform Theor 47(4):1423–1443 12. Chung K-L, Yang W-N, Huang Y-H, Wu S-T, Hsu Y-C (2007) On SVD-based watermarking algorithm. Appl Math Comput 188(1):54–57 13. Cox IJ, Kilian J, Leighton FT, Shamoon T (1997) Secure spread spectrum watermarking for multimedia. IEEE Trans Imag Process 6(12):1673–1687 14. Das C, Panigrahi S, Sharma VK, Mahapatra KK (2014) A novel blind robust image watermarking in DCT domain using inter-block coefficient correlation. AEU - Int J Electron Communicat 68(3):244–253 15. Deb K, Al-Seraj MS, Hoque MM et al. (2012) Combined DWT-DCT based digital image watermarking technique for copyright protection. In: 7th Int Conf Elect Comput Eng 458–461

Multimed Tools Appl 16. Dowling J, Planitz B, Maeder A, Du J, Pham B, Boyd C, Chen S, Bradley A, Crozier S (2008) A comparison of DCT and DWT block based watermarking on medical image quality. In: Shi Y, Kim H-J, Katzenbeisser S (eds) Digital watermarking, vol 5041. lecture notes in computer science. Springer, Berlin Heidelberg, pp 454–466 17. Fan M-Q, Wang H-X, Li S-K (2008) Restudy on SVD-based watermarking scheme. Appl Math Comput 203(2):926–930 18. Feng LP, Zheng LB, Cao P et al. (2010) A DWT-DCT based blind watermarking algorithm for copyright protection. In: 3rd IEEE Int Conf Comput Sci Inform Technol (ICCSIT) 455–458 19. Gunjal BL, Mali SN (2011) Secured color image watermarking technique in DWT-DCT domain. Int J Comput Sci, Eng Inform Technol 1(3):36–44 20. Hernandez JR, Amado M, Perez-Gonzalez F (2000) DCT-domain watermarking techniques for still images: detector performance analysis and a new structure. IEEE Trans Imag Process 9(1): 55–68 21. Hsieh M-S, Tseng D-C, Huang Y-H (2001) Hiding digital watermarks using multiresolution wavelet transform. IEEE Trans Indust Electron 48(5):875–882 22. Kalra GS, Talwar R, Sadawarti H et al. (2014) Adaptive digital image watermarking for color images in frequency domain. Multimed Tools Appl:1–21 23. Laskar RH, Choudhury M, Chakraborty K, Chakraborty S (2011) A joint DWT-DCT based robust digital watermarking algorithm for ownership verification of digital images. In: Venugopal KR, Patnaik LM (eds) Computer networks and intelligent computing, vol 157. communications in computer and information science. Springer, Berlin Heidelberg, pp 482–491 24. Lin SD, Chen C-F (2000) A robust DCT-based watermarking for copyright protection. IEEE Trans Consumer Electron 46(3):415–421 25. Lin SD, Shie S-C, Guo JY (2010) Improving the robustness of DCT-based image watermarking against JPEG compression. Comput Standards Interfaces 32(1–2):54–60 26. Makbol NM, Khoo BE (2014) A new robust and secure digital image watermarking scheme based on the integer wavelet transform and singular value decomposition. Digit Sign Process 33:134–147 27. Mohammad A (2012) A new digital image watermarking scheme based on Schur decomposition. Multimed Tools Appl 59(3):851–883 28. Naderahmadian Y, Hosseini-Khayat S (2014) Fast and robust watermarking in still images based on QR decomposition. Multimed Tools Appl 72(3):2597–2618 29. Patra JC, Phua JE, Bornand C (2010) A novel DCT domain CRT-based watermarking scheme for image authentication surviving JPEG compression. Digit Sign Process 20(6):1597–1611 30. Stanković RS, Falkowski BJ (2003) The Haar wavelet transform: its status and achievements. Comput Elect Eng 29(1):25–44 31. Su Q, Niu Y, Wang G, Jia S, Yue J (2014) Color image blind watermarking scheme based on QR decomposition. Signal Process 94:219–235 32. Su Q, Niu Y, Zhao Y, Pang S, Liu X (2013) A dual color images watermarking scheme based on the optimized compensation of singular value decomposition. AEU - Int J Electron Communicat 67(8):652–664 33. Tao P, Eskicioglu AM (2006) An adaptive method for image recovery in the DFT domain. J Multimed:36–45 34. Tsui TK, Zhang X-P, Androutsos D (2008) Color image watermarking using multidimensional fourier transforms. IEEE Trans Inform Forensics Security 3(1):16–28 35. Vetterli M (1985) Fast 2-D discrete cosine transform. In: IEEE Int Conf Acoustics, Speech, Sign Process 1538–1541 36. Wang Y, Doherty JF, Van Dyck RE (2002) A wavelet-based watermarking algorithm for ownership verification of digital images. IEEE Trans Imag Process 11(2):77–88 37. Wang Y, Pearmain A (2004) Blind image data hiding based on self reference. Pattern Recogn Lett 25(15): 1681–1689 38. Wei ZH, Qin P, Fu YQ (1998) Perceptual digital watermark of images using wavelet transform. IEEE Trans Consumer Electron 44(4):1267–1272 39. Zhang G, Wang S, Wen Q et al. (2004) An adaptive block-based blind watermarking algorithm. In: 7th Int Conf Sign Process (ICSP) 2294–2297 40. Zhao X, Ho AS (2010) An introduction to robust transform based image watermarking techniques. In: Sencar H, Velastin S, Nikolaidis N, Lian S (eds) Intelligent multimedia analysis for security applications, vol 282. studies in computational intelligence. Springer, Berlin Heidelberg, pp 337–364 41. Zheng D, Liu Y, Zhao J et al. (2006) A survey of RST invariant image watermarking algorithms. In: Canadian Conf Electrical Comput Eng Ottawa, Canada 2086–2089

Multimed Tools Appl

Hwai-Tsu Hu received his B.S. degree from National Cheng Kung University, Taiwan, in 1985, and both M.S. and Ph.D. degrees from the University of Florida, USA in 1990 and 1993 respectively, all in Electrical Engineering. Since 1998, he has been a Professor in the Department of Electronic Engineering at National ILan University, Taiwan. His research interests include speech, audio and image signal processing.

Ling-Yuan Hsu received the M.S. degree from National Dong Hwa University, Hualien, Taiwan, in 2004, and Ph.D. degree from the Department of Computer Science and Information Engineering at National Taiwan University of Science and Technology, Taipei, Taiwan, in 2013. He is presently an assistant professor in the Department of Information Management at St. Mary’s Junior College of Medicine, Nursing and Management, Taiwan. His research interests include artificial intelligence, signal processing and evolutionary computation.