1. Introduction
Digital watermarking is a new technology of information hiding and it has effective effect for Copyright Protection. The conception of digital watermarking appeared firstly in the conference of image processing in 1994. Many experts made a lot of research for digital watermarking based on spatial [1] and transform domain [2,3].
From the view of image processing, embedding watermark information can be seemed that a puny signal overlap in a strength signal. Human will not feel the existence of information if the embed strength is lower than the threshold of HVS [4,5].
Robustness of watermarking is important for many applications. Compared with spatial domain methods, transform domain methods have stronger robust. Watermarking based on DWT [6] is an important transform domain method. Compared with the Discrete Cosine Transform (DCT), DWT has a better performance. Firstly, it has good time-frequency decomposition in accordance with the characteristic of HVS. Moreover, DWT has good compatibility for JPEG compression. Embedding watermark information in the DWT domain [7] has better robust against JPEG compression.
In this paper, an algorithm for data hiding by using DWT and JND model of HVS is proposed. It is a dynamic multiple blind watermarking algorithm, since more than one watermarks can be embedded into the original image. Original image is divided into 8*8 blocks and each block is transformed by DWT. A visual model [8] is used to calculate JND threshold. Watermark can be embedded by changing the LL-coefficient of original image.
The paper is organized as follows: Related work is introduced in Section 2. In Section 3 the proposed watermarking algorithm is discussed in detail. Section 4 is experimental results and at last conclusions are drawn.
2. Related Work
JND is defined as a measure referring to the capability of a human observer to detect noise or distortion in the field of view. It is related to characteristics of HVS and adaptive to contents of visual signal under consideration. With a HVS model, accessory information can be embedded in the regions with the least HVS sensitivity to the incurred changes. JND model in the essay of Podilchuk [9] is used. Original image of size M × N is first divided into 8 × 8 blocks (Bm) and each block is decomposed by DWT (
). The formula is as follow:
(1)
(2)
Contrast sensitivity can be calculated in DWT domains by the formula following:
(3)
where u, v is frequency. Noise sensitive visual can be calculated by using Contrast Sensitivity. The formula is as follow:
(4)
where FDWT is DWT of each block. JND can be calculated by:
(5)
where 
is strength which can be obtained by experiment.
3. Proposed Watermarking Algorithm
Original image is decomposed into four sub-band images by DWT. Approximate sub-image (LL) is stable since it contains the majority of image energy. In order to gaining a better robustness, the watermark is embedded into approximate sub-image.
3.1. Watermark Embedding Process
By altering coefficients of approximate sub-image blocks, watermark is embedded. Watermark embedding is implemented through the following procedures.
Step 1: Generate watermark In our scheme, two binary images are used as original watermarks W. It is transformed into 
by Arnold transform. 
is scanned on-line and then transformed into one-dimensional sequence 
, where K is the number of watermark sequence. Two watermarks are considered for embedded labeled as watermark1 
and watermark 2
.
Step 2: Divide original image Original image of size M × N is divided into 8 × 8 blocks which are scanned into one-dimensional horizontal blocks (Bk).
Step 3: Calculate JND
JND calculation for the coefficient in the low-frequency domains can be finished according to the Formula (5).
Step 4: Generate number sequence Put a secure key X into a random number generator to produce random number sequence I. 
, if i = k, then I(i) = I(k).
Step 5: choose image blocks to be embedded Transform each block into DWT coefficient matrix block
. Sub-block 
is selected to embed 
and 
is selected to embed
.
Step 6: Embed Watermarking Let 
denotes LL sub-band coefficient of sub-block
, Coefficients 
and 
are selected to be embedded. Embedding rule is:
If 
and
, exchange coefficients of 
and
. Else if 
and
, exchange coefficients of 
and
.
Embedding is done as Equations (6) and (7):
If 
and
(6)
where
.
If 
and 
(7)
where
.
Where 
is JND value of
. 
is watermarking signals.
Step 7: Obtain watermarked image Inverse DWT(IDWT) is applied to obtain the watermarked image
.
Watermark embedding procedure is shown in Figure 1.
Step 8: Dynamic watermark embedding By using the same watermark Embedding process as Step 1 - Step 7, more than one watermarks can be embedded into original image at any time, which can achieve the dynamic watermarking embedding.
3.2. Watermark Extraction Process
Watermark can be extracted blindly without original image. The steps are as follows.
Step 1: Apply DWT Divide watermarked image 
into 8 × 8 sub-blocks. Each block is scanned and sorted as embedding process. DWT is done to obtain decomposed sub-block
.
Step 2: Regenerate number sequences Regenerate number sequences using the same key to select
.
Step 3: Extract Watermark Let 
denotes LL sub-band coefficient of subblock
. Extract the one-dimensional sequence 
as:
(8)
Step 4: Obtain the perfect watermarks Transform 
into two-dimensional watermark image. The perfect watermarks are obtained after Arnold transform.
Watermark Extraction process procedure is shown in Figure 2.
Figure 1. Watermark embedding procedure.
Figure 2. Watermark embedding procedure.
4. Experimental Results
Experimental results are simulated by programming in Matlab. Different cover images are used as original images and two binary watermark images for the test. Experimental results with 256 × 256 of original Lena image are given. Original image and watermark images are shown in the Figure 3. Embedded and extracted watermarks are shown in the Figure 4.
4.1. Robustness
This section examines the robustness against some attacks. Normative Correlation (NC) is employed to measure the similarity of extracted and original watermark defined as:
(9)
where W is original watermark and 
is extracted watermark with size M × N. The bigger NC is, the better similarity between W and 
is.
Performance analysis results are cited in Table 1. The proposed method is compared to the method of Reference [10] based on DCT and HVS. From the Table 1 we see that the proposed method performs well against JPEG compression with quality factor of 40, the quality
(a)
(b)
(c)
Figure 3. (a) Original image; (b) Watermark 1; (c) Watermark 2.
(a)
(b)
(c)
Figure 4. (a) Watermarked image; (b) Extracted watermark 1; (c) Extracted watermark 2.
of the extracted watermarks are good, this performance is better than the method in [10]. To demonstrate proposed method resistance to noise, gaussian niose was added. The experimental results in Table 1 show that the algorithm can resist this noise attack very well. It can also be
Table 1. The performance results for the proposed method and the method in [10].
observed from Table 1 that the proposed method outperforms in many cases. The proposed method provides better results in painting, line removal, histogram equalization, median filtering and cropping attack.
4.2. Invisibility
Peak signal noise ratio (PSNR) is used to evaluate the invisibility which is given in Equation (10):
(10)
where 
and 
stand for the pixel values of
Table 2. PSNR values of proposed method with different original image (db).
Table 3. Comparison between the proposed scheme and the scheme of reference [10].
original and watermarked image respectively. M and N are height and width of image.
Table 2 shows PSNR of the proposed scheme for different watermarked images. Table 3 represents PSNR of watermarked attacked images and NC values of extracted watermark1 and extracted watermark 2 between the proposed method and the method of Reference [10].