A Semi-fragile Image Watermarking Scheme Exploiting BTC Quantization Data

Abstract

This paper proposes a novel blind image watermarking scheme exploiting Block Truncation Coding (BTC). Most of existing BTC-based watermarking or data hiding methods embed information in BTC compressed images by modifying the BTC encoding stage or BTC-compressed data, resulting in watermarked images with bad quality. Other than existing BTC-based watermarking schemes, our scheme does not really perform the BTC compression on images during the embedding process but uses the parity of BTC quantization data to guide the watermark embedding and extraction processes. In our scheme, we use a binary image as the original watermark. During the embedding process, the original cover image is first partitioned into non-overlapping $4{\times}4$ blocks. Then, BTC is performed on each block to obtain its BTC quantized high mean and low mean. According to the parity of high mean and the parity of low mean, two watermark bits are embedded in each block by modifying the pixel values in the block to make sure that the parity of high mean and the parity of low mean in the modified block are equal to the two watermark bits. During the extraction process, BTC is first performed on each block to obtain its high mean and low mean. By checking the parity of high mean and the parity of low mean, we can extract the two watermark bits in each block. The experimental results show that the proposed watermarking method is fragile to most image processing operations and various kinds of attacks while preserving the invisibility very well, thus the proposed scheme can be used for image authentication.

1. Introduction

Over the last two decades, information hiding has become an emerging technology that embeds secret information in image [1-3], video [4-7] or audio [8,9]. In the digital information network age, the possible applications of information hiding technologies become broader and broader, and its important branches is the digital watermark. The first application coming to mind is copyright protection of digital media. In the digital world, it is possible for almost anyone to duplicate or manipulate digital work without degrading the quality. The embedded watermark permits identification of the owner of the work. The second application is maybe in the field of data security, where watermarks can be used for certification, authentication, and conditional access. For example, certification is an important issue for official documents, and we can hide the identity number that is written in clear text on the card as a digital watermark in the identity photo; therefore switching or manipulating the identity photo will be detected. The third application is the authentication of image content. The goal of this application is to detect any alterations and modifications in an image. The paper focuses on this application.

According to the robustness together with the application of the watermarks, digital watermarking methods can be categorized into three classes: robust, fragile and multipurpose approaches. Robust watermarking approaches [1,2,5-7,10-12] are mainly designed for copyright protection where the watermark is still detectable after accidental or malicious attacks, while fragile watermarking schemes [3,8,13,14] are basically designed for integrity verification and content authentication where the slightest alteration of the image is detectable in the extracted watermark. Recently, several multipurpose watermarking methods [4,15,16] have been proposed to simultaneously achieve multiple purposes such as copyright protection, content authentication, image retrieval and data hiding. And the semi-fragile watermarking is robust to the operation on image which keeps the image content, such as image compression and image enhancement, etc. while it is fragile for the malicious modification. So it is a multipurpose approach.

In the past two decades, since more and more images have been stored in compressed formats such as JPEG and JPEG2000 or transmitted based on vector quantization (VQ) and block truncation coding (BTC), more and more scholars have been engaged in the compressed-domain watermarking schemes. Among them, VQ-based and BTC-based methods are attractive for they are two famous block-based image compression techniques with easy implementation and high efficiency. In the past ten years, several VQ based watermarking approaches [10,11,13,15] have been proposed as a special branch in the digital watermarking area, where the watermark information is embedded in codeword indices. They can be divided into three categories: robust, semi-fragile and multipurpose schemes. The algorithms proposed in [10,11] are robust, the method presented in [13] is semi-fragile, and the scheme provided in [15] is a multipurpose scheme for both copyright protection and content authentication.

In addition to VQ, BTC [17] is another famous block-based lossy image compression technique. It uses a one-bit quantizer to reduce the number of gray levels in each block while preserving the same mean and standard deviation. BTC is a quick, effective and simple block-based lossy image compression technique. It has some characters such as high performance and high channel fault tolerance. And it has a great application value on real-time image transmission. Among several BTC variations, the absolute moment BTC (AMBTC) [18] preserves the first absolute moment instead of the standard deviation along with the mean. AMBTC is computationally simpler than the original BTC. In the past decade, several watermarking and data hiding schemes for BTC compressed gray-level images have been proposed. The first work was proposed by Lu et al. [12], where the robust watermark is embedded by controlling the VQ-BTC encoding process according to the watermark bits. Later, Lin and Chang [19] put forward a data hiding approach for BTC compressed images by implementing LSB substitution operations on BTC high and low means as well as performing the minimum distortion algorithm on BTC bitplanes. Chuang and Chang [20] presented a hiding algorithm to embed data in the BTC bitplanes of smooth blocks. In addition, Hong et al. [21], Chen et al. [22], and Zhang et al. [23] have proposed several lossless data hiding schemes for BTC-compressed images. Recently, Yang and Lu [14] have proposed a fragile image watermarking scheme whose main idea is to exploit VQ or BTC to encode each block according to the corresponding watermark bit.

The previous BTC-based watermarking or data hiding approaches modify either the BTC encoding process or the BTC-compressed data according to the secret bits, and thus the quality of the watermarked image is the same as or worse than that of the BTC-compressed image. Especially, there are few BTC-based fragile image watermarking schemes that can identify the modified areas and recover the original content. To improve the watermarked image quality and not be detected by the malicious attackers, this paper proposes a new approach that does not really perform the BTC compression on the image but uses the AMBTC quantized data to guide the embedding and extraction processes, and thus lots of data can be embedded as well as a high image quality can be obtained. In addition, the proposed scheme can locate the maliciously modified locations. The rest of this paper is organized as follows. First, Section 2 briefly introduces the AMBTC technique followed by five related works to be compared in the experiments. Section 3 provides a detailed description of the proposed algorithm. Experimental results and a comparison with existing seven works are given in Section 4. Finally, Section 5 summarizes the contributions of our work.

 

2. Related Work

2.1. Absolute Moment Block Truncation Coding

The absolute moment BTC (AMBTC) [18] is a block-based spatial domain image compression technique. The main idea is to quantize every pixel in an image block into two levels while preserving the mean and the first absolute central moment of a block. The AMBTC method first segments the 256-graylevel input image X into non-overlapping small blocks, each being a k-dimensional vector (typically k=4×4), namely, X={x1, x2, …, xN}, where N is the number of blocks. Then, for each block xi=(xi1, xi2, …, xik)T, the AMBTC scheme separately quantizes each element in xi into two levels such that the mean value and the first absolute central moment can be preserved in the reconstruction stage. The mean value mi of each block xi is taken as the one-bit quantizer’s threshold, namely,

Let qi stand for the number of pixels having value not less than mi. Then the two output quantization level values ai and bi can be calculated as follows

where ai and bi are defined as the low mean and the high mean of block xi, respectively. If qi=k, we define ai=bi=mi. After the binary quantization is performed for all the pixels in xi, we can obtain a bitplane pi in which ‘0’ corresponds to the pixels with values less than mi while ‘1’ corresponds to the rest of the pixels. Thus we can use the triple (ai, bi, pi) to describe the compressed version of xi. During the decoding stage, the image block xi can be easily reconstructed from the bitplane pi by replacing each ‘0’ with ai and each ‘1’ with bi respectively. Fig. 1 shows an example of encoding and decoding the image block x based on AMBTC. Obviously, for a 256-graylevel 4×4-sized image block, its low mean and high mean are coded separately with 8 bits each, and its bitplane needs 16 bits, so the coding bit rate is (8+8+16)/16=2bpp. Although the coding bit rate is much higher than JPEG and VQ, AMBTC can be performed very fast.

Fig. 1.An example of encoding a block x by the triple (a, b, p).

2.2. Chuang and Chang’s Scheme

For a smooth block xi whose absolute distance between ai and bi is small, its bitplane pi will be less significant in the AMBTC reconstruction process. Under these circumstances, some suitable locations in pi may be replaced with the s2.3. Hong et al.’s Schemeecret bit. Based on this idea, Chuang and Chang’s scheme [20] embeds data in the smooth blocks’ bitplanes that are determined by the difference between bi and ai. If bi-ai is not greater than the preset threshold TH, then the block xi is classified as a smooth block; otherwise, the block is a complex block. Then, a suitable location in the smooth block’s bitplane pi is selected to be replaced with the secret bit. Obviously, the higher the threshold TH is, the more data may be hidden, but the more distortion will be introduced. The extraction process is very simple. From each compressed BTC blocks, the difference bi-ai is first calculated, and then whether the difference is less than TH or not is determined. Once confirmed, the secret bit in the bitplane pi is extracted. One obvious drawback of Chuang and Chang’s scheme is that the capacity is determined by the number of smooth blocks.

2.3. Hong et al.’s Scheme

As we know, if we want to exchange ai and bi in the compressed triple of block xi, we only need to flip the bitplane pi into in order to obtain the same reconstructed block. Based on this idea, Hong et al.’s embedding scheme [21] can be illustrated as follows: First, all embeddable blocks with ai bi, then the secret bit is 1; if ai

2.4. Chen et al.’s Scheme

To embed secret data in the blocks with ai=bi, Chen et al. [22] proposed an improved reversible data hiding algorithm by introducing Chuang and Chang’s bitplane replacement idea to deal with the case ai=bi in Hong et al.’s bitplane flipping scheme. The scheme aims to get a payload higher than Hong et al.’s bitmap flipping scheme as well as a stego-image quality improvement over Chuang and Chang’s scheme.

2.5. Yang and Lu’s Method

The above three methods fall into the data hiding category that is a bit different from digital watermarking. Recently, Yang and Lu [14] have proposed a very simple BTC/VQ-domain watermarking scheme. Their embedding procedure can be detailed as follows: the original image X is first divided into non-overlapping blocks xi, i=1, 2, …, N. For each image block xi, if the watermark bit to be embedded wi=1, then VQ is used to encode the block xi by searching for its best-matched codeword cj in the codebook C. If the watermark bit to be embedded wi=0, then the AMBTC is used to encode the block xi by replacing the pixels that are not less than the mean value mi with the high mean bi and replacing the pixels that are smaller than the mean value mi with the low mean ai. Finally, all encoded image blocks are pieced together to obtain the final watermarked image Xw. The extraction process is just the reverse process of the embedding process as follows: the suspicious image Y is first divided into non-overlapping blocks yi, i=1, 2, …, N. For each image block yi, it is encoded with VQ and BTC to obtain their corresponding mean squared error MSEVQ and MSEBTC respectively. If MSEVQ < MSEBTC, then the extracted watermark wi=1. If MSEVQ > MSEBTC, then the extracted watermark bit wi=0. If MSEVQ = MSEBTC, then wi is randomly set to 0 or 1. All obtained watermark bits are pieced together to obtain the final extracted watermark We.

2.6. Zhang et al.’s Scheme

Zhang et al. [23] have proposed an oblivious fragile watermarking scheme for images utilizing edge transitions in BTC bitmaps. In the embedding process of the scheme, the original image was partitioned into non-overlapping 4×4 blocks, then the bitmap of each block was been obtained with BTC. The embedding watermark bit was been used to guide to modify the pixel values until the parity of the number of horizontal edge transitions in the bitmap of the modified block is equal to the embedding watermark bit. Similarly, in the extraction process, only to check the parity of the number of horizontal edge transitions in the BTC bitmap of each non-overlapping 4×4 blocks. Then one watermark bit would be extracted from each block.

 

3. Proposed Watermarking Scheme

3.1. Preprocessing Stage

From the above discussion, we can see that the existing BTC-based information hiding schemes actually embed the information in the BTC-compressed image. Thus, the watermarked image is usually with a poor quality. In order to improve the image quality, this paper proposes a new train of thought. We still exploit the BTC, but we do not perform the watermark embedding on the BTC compressed image but use the BTC quantization data to guide the watermark embedding and extraction process. The proposed algorithm consists of three processes, namely, the preprocessing stage, the watermark embedding stage and the watermark extraction stage, as shown in Fig. 2.

Fig. 2.The block diagram of the proposed watermarking scheme.

From Fig. 2, we can see that the preprocessing stage is the common stage for both watermark embedding and extraction processes. The purpose of this stage is to achieve the BTC quantization data as the control parameter. Assume the input image is X or Y consisting of N blocks and the watermark is a binary image W with 2N bits. Thus, the preprocessing stage is just a BTC coding process as follows:

Step 1: The input image X (or Y) is divided into non-overlapping k-dimensional blocks xi (or yi), i = 1,2,…N.

Step 2: Each image block xi (or yi) is encoded by AMBTC, obtaining its mean value mi and its two quantization levels ai and bi.

Step 3: All means are composed of the mean sequence M = {mi|i = 1,2,…,N}, all low means are composed of the low mean sequence A = {ai|i = 1,2,…,N} high means consist of the high mean sequence B = {bi|i = 1,2,…,N}

3.2. Watermark Embedding Stage

With the means M = {mi|i = 1,2,…,N} the low means A = {ai|i = 1,2,…,N} and high means B = {bi|i = 1,2,…,N} in hand as the guider, now we can describe our watermark embedding process. Before embedding, we perform the raster scanning on the original watermark W , obtaining the watermark bit sequence W = {w1,w2,…,W2N}. The purpose of our embedding process is to ensure the blind extraction such that the watermark can be extracted from the watermarked image only based on its two quantization values ai and bi.

Here, a function Parity(x) is defined as follows:

The central idea is to force the parity of ai and the parity of bi to be equal to the two watermark bits. That is, two watermark bits will be embedded in each block. The embedding process can be illustrated detailedly as follows:

Step 1: The original image X is divided into non-overlapping k-dimensional blocks xi, i = 1,2,…,N.

Step 2: The embedding process is performed block by block. For each block xi, the two watermark bits to be embedded are w2i-1 and w2i, the embedding procedure is controlled by the quantization data ai and bi as following substeps.

Step 3: If every block has been embedded with 2 watermark bits, then the algorithm is terminated with the watermarked image Xw. Otherwise, go to Substep 2.1 for next block.

In order to understand the proposed embedding process more clearly, two concrete examples are described in Fig. 3. For the uniform block x1, assume the two watermark bits to be embedded are w1=1 and w2=0, since a1=b1=2, thus Parity(a1)≠w1 and Parity(b1)=w2, and it falls into Sub-case 1.1. For block x2, assume the two watermark bits to be embedded are w3=1 and w4=1, since a2=4 and b2=11, thus Parity(a2)≠w3 and Parity(b2) =w4, and it falls into Sub-case 2.1.

Fig. 3.Concrete embedding examples.

3.3. Watermark Extraction Stage

According to Fig. 2 (b) , we can see that our watermark extraction process is a blind process such that we do not require the original image during the extraction process. It is very simple because we can get the two watermark bits from each block only based on checking the parities of ai and bi. Before extraction, the same preprocessing step is performed on the suspicious image Y to obtain its low mean sequence A={ai |i=1,2,…,N} and high mean sequence B={bi |i=1,2,…,N}. For each block yi, if ai is odd, then we can extract the watermark bit w2i-1=1. Otherwise, we can extract the watermark bit w2i-1=0. Similarly, if bi is odd, then we can extract the watermark bit w2i=1. Otherwise, we can extract the watermark bit w2i=0. After all blocks are performed, we piece these bits together to obtain the final extracted watermark We={we1, we2,…, we2N}.

 

4. Experimental Results and Analysis

4.1. Performance Testing

To evaluate the performance of the proposed watermarking scheme, the 256 grayscale 512×512 sized Lena image is used for watermarking. The Lena image is divided into 16384 blocks of size 4×4. A binary image of size 256×128 serves as the watermark W for embedding. Fig. 4(a) shows the original Lena image and Fig. 4(b) shows the watermarked Lena image with PSNR=51.11dB. To check the fragility of our algorithm, we perform several attacks on the watermarked image, including brightness enhancement by 10%, contrast enhancement by 10%, JPEG compression with QF=100%, JPEG compression with QF=80%, image cropping in the middle part of the image, blurring with a 3×3 window and a threshold 25, median filtering with a 3×3 window, and rotation by 1˚. The attacked images are shown in Fig. 5(a)-(h). The corresponding extracted watermarks are shown in Fig. 6(c)-(j), while the original watermark is shown in Fig. 6(a) and the extracted watermark under no attacks is shown in Fig. 6(b). The performance of extracted watermarks are evaluated by normalized cross-correlation (NC). From these results we can see that the proposed algorithm is fragile to most attacks except for the change in brightness which is caused by the AMBTC’s special ability in preserving the mean value and the first absolute central moment. In that the texture of the image belongs to the part of the ownership and the operation on brightness enhancement can not change the texture of the image, so our proposed scheme can be used for texture authentication. From Fig. 5(e) and Fig. 6(g), we can see that our scheme can locate the maliciously modified locations. In addition, the proposed algorithm is blind because the original image is not required during the extraction process.

Fig. 4.The cover and watermarked images.

Fig. 5.The attacked watermarked images under 8 attacks, namely, brightness enhancement, contrast enhancement, JPEG (QF=100), JPEG (QF=80), cropping, blurring, median filtering, and rotation, respectively.

Fig. 6.The original watermark, the extracted watermark from unattacked watermarked image and extracted watermarks under various attacks, namely, brightness enhancement, contrast enhancement, JPEG (QF=100), JPEG (QF=80), cropping, blurring, median filtering, and rotation, respectively.

In addition, we have two methods to improve the security of our proposed scheme. One method is to keep the size of the non-overlapping k-dimensional blocks as a key. The other method is to select the different position from the image as the first block’s left-top pixel, and the different position is a key.

4.2. Performance Comparison

To show the superiority of our proposed algorithm, we use six test images, Lena, Peppers, Mandrill, Boat, Goldhill and Jet_F16, of the same size 512×512 with 256 grayscales, as shown in Fig. 7. Comparisons among our algorithm, Chuang and Chang’s algorithm, Hong et al.’s algorithm, Chen et al.’s scheme, Yang and Lu’s Method and Zhang et al.’s scheme which are described in Section 2 are performed under the same block size 4×4. Two aspects of performance are considered, namely, the peak signal to noise ratio (PSNR) representing the quality of the watermarked image, and the capacity representing the maximum number of secret bits that can be hidden. Here, although the capacity is not a very important index for a watermarking scheme, but we would like to demonstrate that our scheme can hide large number of bits while preserving high image quality compared to existing BTC-based schemes.

Fig. 7.Six test images.

As shown in Table 1, the PSNRs of watermarked images based on our scheme are much higher than those of the existing schemes. With respect to the embedding capacity, compared with existing schemes, our scheme can embed 2bits per block, which is at least double capacity of the existing schemes. Taking the above two attributes into comprehensive consideration, the proposed scheme is a better method for its high capacity and high PSNR.

Table 1.Comparisons between the proposed and existing BTC-based hiding schemes (k=4×4)

We also compare our scheme with two classical LSB-based schemes, one is proposed by Mielikainen [24], the other is proposed by Zhang and Wang [25]. In Mielikainen’s scheme, he proposed a modified LSB matching-based steganographic method for embedding message bits into a still image. The embedding operation is performed on every two pixels. The LSB of the first pixel carries one bit of information, and another bit of information is presented by a function of the two pixel values. Their method made fewer changes to the cover images comparing to LSB matching and improved the image quality. However, there was a problem to handle with the saturated pixels, i.e., the pixels with value‘0’ or ‘255’ in the 256-grayscale image. That is, the number of non-saturated pixels is the capacity of this scheme. In Zhang and Wang’s scheme [25], a secret digital in (2n+1)-ary notational system is embedded into an n pixels group, in which only one pixel is increased or decreased by 1. In our experiment, we set n=54, then we have the fixed capacity 32856(=ln109/ln2*512*512/54). The comparison results are shown in Table 2. From this table, we can see that the two LSB-based schemes have better performance than our scheme in terms of PSNR and capacity, for they are pixel-based method and only modify the LSB of the pixel to be embedded, while our scheme is block-based and we modify more than one pixel in each block. However, our scheme is a semi-fragile scheme that is robust to brightness or contrast changes, thus our scheme has some special characteristics that the LSB-based schemes do not have. In addition, we provide a new watermarking mechanism of pixel-domain embedding followed by compressed-domain analysis and extraction.

Table 2.Comparisons between the proposed and recent two LSB-based watermarking schemes

 

5. Conclusion

Considering the existing BTC-based watermarking or data hiding schemes obtain low-quality images or have less hiding capacity, in this paper, we propose a new blind fragile image watermarking approach by exploiting the AMBTC technique to guide the watermark embedding and extraction processes. It is a simple but efficient method suitable for texture authentication and data hiding. The central idea is to force the parity of the low mean and high mean of each block to be equal to the two watermark bits to be embedded respectively. Another contribution of this paper is that we have developed a semi-fragile watermarking scheme with the mechanism of pixel-domain-based embedding and compressed-domain-based extraction, which can make full use of the statistical characteristics of the compressed-domain. Experimental results show that it is fragile to most kinds of attacks and it is better than some existing BTC-based image watermarking schemes. The future work will be concentrated on how to further improve the capacity while maintaining the high image quality and how to recover the maliciously modified blocks.