Image Watermarking Scheme Based on Scale-Invariant Feature Transform

Abstract

In this paper, a robust watermarking scheme is proposed that uses the scale-invariant feature transform (SIFT) algorithm in the discrete wavelet transform (DWT) domain. First, the SIFT feature areas are extracted from the original image. Then, one level DWT is applied on the selected SIFT feature areas. The watermark is embedded by modifying the fractional portion of the horizontal or vertical, high-frequency DWT coefficients. In the watermark extracting phase, the embedded watermark can be directly extracted from the watermarked image without requiring the original cover image. The experimental results showed that the proposed scheme obtains the robustness to both signal processing and geometric attacks. Also, the proposed scheme is superior to some previous schemes in terms of watermark robustness and the visual quality of the watermarked image.

1. Introduction

With the rapid developments in multimedia technology and the Internet over the last decade, digital data, i.e., images, videos, audio, and text, can be easily copied and altered when they are transmitted via the Internet. Hence, protection of the ownership of digital data has become a very essential issue. Many solutions have been proposed to solve this issue, and digital watermarking is one of the most promising. Watermarking schemes can embed specific data, also referred to as a ‘watermark,’ into the original image. The watermark cannot be easily seen, which means that an unauthorized person cannot visually detect that embedded data in the content of the watermarked image. To prove the ownership of the image, the embedded watermark is extracted and detected. In watermarking applications, it is most important that the watermark must be sufficiently robust to withstand different attacks. There are two types of watermark attacks, i.e., signal processing and geometric attacks. Watermarking schemes [1-17,20-25] can be classified into template-based [1,2], invariant transform domain-based [3-5], histogram-based [6,7], moment-based [8,9], and feature-based schemes [10-17].

Many feature-based image watermarking algorithms [10-17] have been introduced in the literature. This is because these algorithms can resist signal processing attacks, and they are robust against geometric attacks. In [13], Bas et al. utilized the Harris detector technique to determine the feature points from the original image. A Delaunay Tessellation technique is implemented to divide the image into a set of disjointed triangles. Then, the watermark is embedded into each triangle of the tessellation. However, in this scheme, the feature points extracted from the original image and from the attacked image are not the same. In other words, their schemes cannot extract exactly the embedded watermark once the watermarked image is attacked. In [11], Li and Guo applied the Harris detector to determine the non-overlapped circular areas and embedded the watermark into the spatial domain of these areas. However, in using the spatial domain for embedding the watermark, their scheme had limited robustness against both signal processing attacks and geometric attacks. In [14,15], Seo and Yoo proposed two watermarking schemes using a multi-scale Harris detector. In their schemes, the image is decomposed into disjointed, local circular regions. In [14], a circular, symmetric watermark was embedded after scale normalization processing according to the local characteristic scale, whereas, in [15], Seo and Yoo extract the selected regions and the watermark is embedded in these regions after geometric normalization processing adopting to the shapes of these regions. In [12], Lee et al. extracted local circular regions by using scale-invariant feature transform (SIFT). In the extracted local circular regions, pixel values are modified to embed the watermark. Because the watermark is embedded in spatial domain in these schemes [11,12,14,15], their embedded watermarks are less robust. To further improve the robustness of watermarking schemes, Wang et al. [17] proposed the feature-based watermarking scheme in the transform domain. Their scheme achieved high robustness of the watermark against signal processing attacks. However, the size of the watermark is quite small, i.e., only32 bits. To enhance the watermark’s robustness and size, Li et al. [16] proposed a new watermarking scheme in the DWT domain. The watermark is first resized to the size of the horizontal and vertical high-frequency DWT subbands of the images. Then, the difference between the horizontal and vertical high-frequency DWT coefficients is expanded to embed the watermark. Their scheme outperformed schemes proposed by Lee et al. [12] and Wang et al. [17].

In this paper, a new, robust, watermarking scheme is proposed. The proposed scheme uses the SIFT algorithm to extract the feature areas for embedding the watermark. To achieve better robustness, the DWT coefficients of the SIFT areas is utilized for carrying the watermark. Our experimental results showed that the proposed scheme is resilient to both signal processing attacks and geometric attacks, e.g., Salt and Pepper noise, Gaussian filtering, rotation, and cropping. In addition, the visual quality of the watermarked image is excellent in the proposed scheme.

The rest of this paper is organized as follows. Section 2 provides the concept of the SIFT algorithm [18] to give readers sufficient background knowledge. Section 3 describes the proposed scheme, consiting of the watermark embedding and extracting phases. Section 4 presents the experimental results and illustrates the superiority of the proposed scheme. Finally, we present our conclusions in Section 5.

 

2. SIFT Algorithm

In 2004, Lowe [18] first introduced the scale-invariant feature transform (SIFT) algorithm, which proved that the extracted feature points are stable to geometric transformations, i.e., scaling, rotation, and translation transformation. The SIFT algorithm extracts the feature points from the scale space of the image. The scale space of image is denoted as L(x, y, α), and is defined in Equation (1):

where I(x, y) is the digital image, * denotes the convolution operation, and Gau(x, y, α) is the variable-scale Gaussian kernel with a standard deviation α. Gau(x, y, α) is defined in Equation (2):

To detect the SIFT feature points, scale-space extrema in the difference-of-Gaussian (DoG) function, D(x, y, α) is applied, and it can be computed using Equation (3):

where k is a constant, multiplicative factor; Gau(x, y, kα) is the variable-scale Gaussian kernel with the standard deviation α and the constant multiplicative factor k; and L(x, y, kα) is the scale space of the image with the constant multiplicative factor k. To determine the candidates of the difference-of-Gaussian function, D(x, y, α), each point is first compared with its eight neighbors points in the current image. Then, it is also compared to the nine neighbors in the scale above and below. The point is selected only when its value is larger or smaller than the value of all of these neighbors. In addition, the positions that have low contrast or are poorly localized are deleted by stability function.

After obtaining the feature points, one or more orientations are assigned to each point on the basis of the local image gradient directions. Let the gradient magnitude be gm(x, y), and orientation of the feature point (x, y) be O(x, y), which are calculated using Equations (4)-(5), respectively.

where L is the Gaussian smoothed image [3]. The descriptor of feature area is constructed by first calculating the gradient magnitude and orientation at each image sample point in an area around the feature point location. Then, these samples are accumulated into orientation histograms that summaries the contents over 4 × 4 sub-areas, with the length of each orientation bin corresponding to the sum of the gradient magnitudes near that direction within the area. The feature area is formed with containing the values of all the orientation histograms entries, corresponding to the lengths of the orientation bins. In the SIFT feature descriptor, the radius of an area around the feature point is determined as a constant times the detection scale of the feature point, which can be motivated by the property of the scale selection mechanism in the feature point detector of returning a characteristic size estimate associated with each feature point [26].The reader can refer to [18] for the detailed algorithms that are used to extract the features of SIFT. In the research reported in this paper, the SIFT algorithm was used to extract the feature areas of the image in which the watermark was embedded.

 

3. The Proposed Schemes

After applying the SIFT algorithm, several SIFT feature areas were determined in an image. However, a problem arises if the SIFT feature areas are generated by using the SIFT algorithm, since some of the feature areas may overlap. In order to hide the watermark, some local areas must be removed. If two areas are overlapped, only the area whose size is larger than or equal to 60 × 60 pixels is reserved. The main reason is that areas of these sizes are most suitable for embedding watermark sizes of 32 × 32 pixels in the proposed scheme. In addition, if two areas larger than 60 × 60 pixels are overlapped, only the one that has the larger DoG function value, calculated by Equation (3), is reserved because the larger DoG value offers the better stability. Fig. 1 shows examples of the extracted SIFT feature areas using SIFT.

Fig. 1.Example of extracted SIFT feature areas from the image Lena

Each extracted SIFT area is processed independently for embedding the watermark. To further improve the robustness of the watermark, N extracted SIFT areas are used to carry the same copy of the watermark. Notably, since all circular SIFT areas larger than 60 × 60 pixels are selected for embedding the watermark, then N is equal to 20 if there are 20 areas having the size larger than 60 × 60 pixels. In addition, watermark embedding and watermark detection are performed in the DWT domain. DWT is a mathematical technique that can be used to transform the image in the spatial domain into the frequency domain. Basically, the DWT image is obtained repeatedly by filtering the current image on a row-by-row and a column-by-column. After each level DWT transformation, four sub-bands, i.e., CAi, CHi, CVi, and CDi, are generated. Fig. 2 shows an example of one-level DWT transformation.

Fig. 2.Example of one-level DWT transformation

In the DWT domain, the CAi sub-band is not used for carrying a watermark. It is because CAi is a low-frequency sub-band, and this sub-band will contain essential information about the image. Therefore, small modifications in this sub-band can easily cause the image to be distorted. Similarly, embedding the watermark into sub-band CDi also is not applied, because this sub-band can be eliminated easily, for example, via JPEG compression. Thus, in the work described in this paper, the CHi and CVi sub-bands were used to carry the watermark. In addition, since the watermark W is partitioned and embedded into both the horizontal and vertical frequency sub-bands, i.e., CHi, and CVi, of the selected SIFT areas. Therefore, the size of watermark W that is used should be the smaller or equal to the half size of the horizontal and vertical frequency sub-bands, i.e., CHi, and CVi, of the selected SIFT areas. The main reason is to guarantee robustness of watermark when the watermark with the original size is completely embedded into these sub-bands.

In this paper, to resist the geometric attacks, i.e. scaling and rotation, the circular feature area is used for watermarking. As a result, scaling invariance can be obtained because the radius of area is directly proportional to the scale of feature area. Moreover, to achieve the rotation invariance, circular area should be required. Fig. 3 shows how pixels present in the circular are and rectangle area. The proposed scheme contains two phases, i.e., the embedding phase and the extracting phase, and these two phases are described in Subsections 3.1 and 3.2, respectively.

Fig. 3.Different shapes of feature areas

3.1 Watermark Embedding Phase

In this subsection, the same watermark is embedded repeatedly into N extracted SIFT areas. Fig. 4 shows the flowchart of watermark embedding phase. Then the watermark embedding algorithm is described.

Fig. 4.Flowchart of watermark embedding phase

Step 1: First, N SIFT areas having its size larger than 60 × 60 pixels are selected from the original image, and the size of the SIFT area is determined by the size of the watermark. In this paper, the size of the SIFT area is larger than 60 × 60 pixels, and the size of the image is 512 × 512 pixels; the size of the watermark can be 32 × 32 pixels.

Step 2: Each selected SIFT area is resized to 64 × 64 pixels, and one-level DWT is applied to yield four frequency sub-bands {CAi, CHi, CVi, CDi}.

Step 3: Partition the watermark image W into two parts, W1 and W2, as shown in Fig. 4.

Step 4: Embed the two parts of the watermark image, W1 and W2, by modifying the coefficients in the horizontal and vertical frequency sub-bands, i.e., CHi, and CVi, respectively, as shown in Fig. 5. It is noticeable that only the DWT coefficients locate inside the left, upper half of the inscribed circle, which is modified for embedding the watermark.

Fig. 5.Demonstration of watermark embedding

In the proposed scheme, W1 is embedded into the horizontal frequency sub-band CHi, and W2 is embedded into the vertical frequency sub-band CVi. To embed watermark bit W1(x, y), the corresponding horizontal, high-frequency coefficient, that has the similar coordinates are selected as CHi(x, y), is used. Then, the two digits Hi of the fractional portion of CHi(x, y) are extracted to embed watermark bit W1(x, y). Notably, the value of Hi is extracted from the value of CHi, and Hi is the two digits that are selected starting from the first non-zero digit of the fractional portion of CHi. For example, if CHi(x, y) = 0.001052, the value of Hi will be 10. Similarly, W2(x, y) is embedded into the two digits Vi that are selected starting from the first non-zero digit of the fractional portion of CVi(x, y). To explain further, let b denote W1(x, y). (or W2 (x, y)) and let p denote the result of Hi mod T (or Vi mod T). Note that Hi (or Vi) is two digits that are selected starting from the first non-zero digit of the fractional portion of CHi(x, y) (or CVi(x, y)), hence, Hi (or Vi) is in range [10, 99]. Notably, the value of Vi is extracted from the value of CVi, and Vi is first two non-zero digits of the fractional portion of CVi. Here, T is the threshold value. Therefore, p = Hi mod T (or Vi mod T) must be in the range [0, T). The watermark bit b is embedded into the value p using Equation (6):

where T is the threshold value; p is the result value of Hi mod T if CHi is used to embed watermark W1(x, y) (or Vi mod T, if CVi is used to embed watermark W2 (x, y)); and b is the watermark bit W1(x, y) (or W2 (x, y)). Fig. 6 shows the value of p and the value of p’ based on the threshold T after embedding watermark bit b.

Fig. 6.Value of p’ after embedding the watermark

Step 5: If CHi is used to embed watermark W1(x, y), set CHi’ to the value of Hi’ = p’. Otherwise, if CVi is used to embed watermark W2(x, y), set CVi’ to the value of Hi’ = p’.

Step 6: Utilize one-level IDWT to reconstruct the watermarked SIFT area. This area is resized to its original size and is used to substitute for the original SIFT area in the image.

Step 7: Repeat Steps 2 through 5 until N SIFT areas have been completely processed.

3.2 Watermark Extracting Phase

This Subsection describes how watermark W is extracted from the watermarked image. The extracting algorithm is implemented without requirement of the original image. Therefore, the proposed scheme meets the condition of blindness. Some of the steps in the watermark extracting algorithm are exactly the same as those used in watermark embedding phase. Specifically, SIFT algorithm is used to extract N watermarked SIFT areas from the watermarked image and resized them to 64 × 64 pixels. Then, the watermark is extracted from each watermarked SIFT area. Fig. 7 shows the watermark extracting processes and the watermark extracting algorithm is performed in each SIFT area, as shown below.

Fig. 7.Watermark extracting procedure

Step 1: Utilize one-level DWT on each watermarked SIFT area to produce four frequency sub-bands {CAi’, CHi’, CVi’, CDi’}.

Step 2: For the coefficients in two sub-bands, i.e., CHi’ or CVi’, the value p’ is determined as first two non-zero digits of the fractional portion of CHi’ or CVi’ to extract watermark bit b’. Note that only coefficients inside the left, upper half of the inscribed circle of sub-bands, CHi’ and CVi’, are processed. In this scenario, if b’ is extracted from CHi’ coefficients (or CVi’ coefficients), it belongs to W1 (or W2), respectively. Watermark bit b’ is extracted using Equation (7).

Step 3: Combine two watermarks W1 and W2 to obtain the extracted watermark W’. Then, based on the demonstration of the extracted watermark, the final decision will be made. In addition, the watermark similarities should be calculated to judge objectively.

 

4. Experimental Results

In this section, the performance of the proposed scheme in terms of the watermark’s being invisible and robustness is demonstrated. The experiments were performed on a standard image with the size of 512 × 512 pixels. The watermark is a circular binary image size of 32 × 32 pixels. N SIFT areas were selected for embedding the watermark. In this experiment, three different values of N, i.e. 3, 4, and 5, are used.

The peak signal-to-noise ratio (PSNR) was used to estimate the visual qualiy of the watermark. The PSNR was computed using Equation (8):

where M is the size of the original image, and Ii and I’i are the pixel values before and after embedding the watermark, respectively. Table 1 lists the PSNRs of the watermark image with different values of threshold T. It is apparent that the proposed scheme provided watermarked images with high visual quality and different values of threshold T. Fig. 8 shows image Lena before and after the watermark was embedded. Fig. 8 clearly shows that the proposed scheme made the watermark invisible when the value of PSNR is greater than 84 dB.

Table 1..PSNR of the watermarked image with different values of threshold T

Fig. 8.Watermark’s invisibility

To demonstrate the robustness of the proposed scheme, we use the normalized correlation coefficient (NC) to measure the similarity between the embedded watermark W and the extracted watermark W'. NC is calculated using Equation (9):

where Wh and Ww are the height and width of the embedded watermark, respectively. Note that, in this experiment, the highest NC value from N SIFT areas is selected for comparisons.

Fig. 9 shows performance of the proposed scheme with several thresholds T, i.e., 20, 30, 40, and 50. To obtain the tradeoff between visual quality of the watermarked image and robustness, the threshold T = 30 was selected in the proposed scheme.

Fig. 9.Performance of the proposed scheme with several thresholds

Signal processing and geometric attacks listed in the benchmark software, Stirmark 4.0 [19], were used on the watermarked images. These attacks try to weaken or remove the embedded watermark from the watermarked images. Fig. 10 shows that the embedded watermark can be extracted clearly from all watermarked images under Salt and Pepper noise, JPEG 100, and Gaussian attacks with different parameters. As can be seen in Fig. 10, the proposed scheme obtained a high value of NC, i.e., greater than 0.9. This indicates that our proposed scheme was resilient against these attacks and that it ensured the high visual quality of the embedded watermarked images, as shown in Table 1.

Fig. 10.Watermarked images, extracted watermarks, and NC values for Salt and Pepper noise, JPEG 100, and Gaussian attacks

Fig. 11 shows the results of the proposed scheme for rotation attacks and cropping attacks. In the rotation attacks, three cases of rotating the watermarked images were simulated, i.e., 2°, 5°, and 10°. In the cropping attacks, three different cases were performed to crop three watermarked images. In the first case, the corner of the watermarked image was cropped by 25%. In the second and the third cases, the outsides of the watermarked images were cropped by 50% and 75%, respectively. Fig. 11 shows that the watermark was extracted completely from the watermarked images.

Fig. 10.Watermarked images, extracted watermarks, and NC values for rotation and cropping attacks

To further validate the proposed scheme, we compared it with two previous schemes [12,16]. These schemes were selected to estimate the proposed scheme because they are similar to the proposed scheme in terms of using invariant feature points to conceal the watermark. Table 2 shows that our proposed scheme obtained much higher PSNR than those of the previous schemes [12,16], even though the proposed scheme and Li et al.’s scheme both used the DWT domain to embed the watermark. However, in [16], Li et al. modified the horizontal and the vertical high-frequency DWT coefficients to hide the watermark bits. Whereas, the proposed scheme only hides the watermark bits into the fractional portion of the horizontal or the vertical high-frequency DWT coefficients. Consequently, more distortions of the watermarked images occurred in Li et al.’s scheme than in our proposed scheme.

Table 2.PSNR of watermarked and original images (dB)

Table 3 compares the robustness of the watermarks for Li et al.’s scheme, Lee et al.’s scheme, and our proposed scheme, and it shows that the proposed scheme had greater robustness. In Lee et al.’s scheme, the pixels in the SIFT regions are modified to embed the watermark. Hence, since more pixels have their values altered in an attack, the extracted watermark will have more distortion. Instead of embedding the watermark in the spatial domain, Li et al.’s scheme embeds watermark into the DWT domain. As a result, their scheme obtained higher robustness than Lee et al.’s scheme. However, in Li et al.’s scheme, the difference between the horizontal and vertical high frequency DWT coefficients was expanded to embed the watermark. Therefore, if the value of one of the horizontal or the vertical high frequency DWT coefficients is modified, the extracted watermark bit will be changed. Conversely, in the proposed scheme, only the value of the horizontal or vertical high frequency DWT coefficient is altered to embed the watermark bit. Table 4 shows the robustness of the proposed scheme with different values of N.

Table 3Comparisons of the robustness of the watermark among Li et al.’s scheme, Lee et al.’s scheme, and the proposed scheme

Table 4.Robustness of the proposed scheme with different values of N

Based on experimental results we can see same results were obtained the same with different Ns. This is because only the highest NC results are selected in our scheme, and the highest results can be obtained when N=3. Therefore, we set N as 3 here.

 

5. Conclusions

In this paper, a new image watermarking scheme was proposed to obtain high visual quality of the watermarked image and resistance against various attacks. In the proposed scheme, the SIFT feature areas are extracted, and a watermark image in binary form is embedded repeatedly into the horizontal and vertical high-frequency DWT coefficients of the selected SIFT feature areas. The experimental results showed that the proposed scheme is resilient to various attacks, i.e., signal processing and geometric attacks. In addition, the proposed scheme achieved better visual quality of the watermarked image and stronger robustness to various attacks than some previous schemes. It can be concluded that our watermarking scheme satisfied the demand of real-time applications. In some attacks, i.e. Median filter and Shearing, the robustness of the proposed scheme is not truly high. Therefore, to achieve better robustness to these attacks can be the future work.