This paper presents a novel watermarking scheme for authentication of digital color images in social networks. The procedure consists of the embedding of a binary watermark image, containing the owner information, into the image to be authenticated. In order to minimize the artifacts in the host image the process is carried out in the wavelets domain. Concretely, the watermark embedding is performed in the HL4 and LH4 sub-band coefficients of the red, green and blue channels of the original image, based on an optimal channel selection quantization technique. To ensure a high robustness to tampering and malicious attacks a key-based pixel shuffling mechanism is further used. The reverse process is likewise identified for the extraction of the watermark from the authenticated image. Both embedding and extraction procedures are benchmarked on diverse color images and under the effects of different types of attacks, including geometric, non-geometric, and JPEG compression transformations. The proposed scheme proves to support imperceptible watermarking, while also showing a high resiliency to common image processing operations.
3. UC Lab Kyung Hee University, South Korea 3
• Millions of images are shared every day through the SNS
• Many of these images end up in the hands of unknown
people that use them in an illegal and malicious manner
• Mechanisms for ensuring the ownership and protecting the
copyright are utterly required to settle potential disputes
Image
Watermarking
4. UC Lab Kyung Hee University, South Korea 4
Convenience
Imperceptibility
Robustness
The information should be extracted from the original image
A watermark has to be imperceptible
A watermark needs to be robust against image modifications
5. visualization
frequency domain
• Quality
Optimal color channel selection
• Accuracy rate in the extraction process
Optimal threshold based on the Otsu method.
UC Lab Kyung Hee University, South Korea 5
6. UC Lab Kyung Hee University, South Korea 6
Article Key concept Limitations
Xiang-yang [2012] Fourier transform
Least square support vector
machine (LS-SVM)
High computation time for LS-SVM training.
Niu [2011] Nonsubsampled coutourlet
transform (NSCT)
Support vector regression (SVR)
Low performing NSCT and computation time in
extraction process.
Song [2012] Curvelet transform
Coefficient quantization technique
Weakness under lossy JPEG compression
Chou [2010] Wavelet transform
Just noticeable color difference
(JNCD)
Weakness under geometric operations and hue
modification.
7. UC Lab Kyung Hee University, South Korea 7
Color
Images
4-DWT
Coeffient
Blocking
Coeffient
Difference
Optimum
Selection
Embedding
Rule
Coefficient
Unblocking
4-IDWT
Embedded
Image
Binary
Watermark
Bit shuffling
Combined
Key
Recovered
Watermark
Bit
Reshuffling
Extraction
Rule
Coefficient
Difference
Coefficient
Blocking
4-DWT
Modified
Image
Adaptive
threshold
Otsu method
Channel
Attacks
Embedding process
Extraction process
8. UC Lab Kyung Hee University, South Korea 8
Color
Images
4-DWT
Coefficient
Blocking
Coefficient
Difference
Optimum
Selection
Embedding
Rule
Coefficient
Unblocking
4-IDWT
Embedded
Image
Binary
Watermark
Bit shuffling
Combined
Key
HL4
(C(HL,i))
LH4
(C(LH,i))
Block 1
C(HL,1 )
Block 2 Block n
C(HL,2) C(LH,1 ) C(LH,2)C(HL,n) C(HL,n)
Coefficient difference
Numberofblocks
Before embedding
The embedding process
Coefficient difference
Numberofblocks
After embedding
0-bits
1-bits
y1 y2
Most data on the Internet including images, videos have been not certified and protected to against the unauthorized copy issue.
Therefore, we need a watermarking technique to assign a copyright on our images.
In which a watermark is embedded into a host image and able to extract exactly in contention.
The motivations for proposal of a watermarking method are robustness, imperceptibility and convenience.
-> Explain each of them.
Describe content on the slide
The input of our method is a color image and a watermark is a binary image.
The uniqueness is an optimum color channel selection to apply the embedding rule to achive an imperceptibility at the output.
The wavelet coefficients have been modified based on the value of coefficient difference and watermark bits.
The embedded images in transmission, storage process can be attacks by common digital signal proceses.
In extraction, the watermark bits are extracted based on comparing the threshold with coefficient different value.
The watermark bits after extracted will be reshuffling to obtain the original one.
Describe content on the slide
This slide show the result of image quality after embedding process based on CPNSR and SSIM parameter.
Higher value of CPNSR is better. Different images with different structure and texture have different values of CPSNR.
SSIM should reach to 1.
Some results of watermark robustness after extraction under various attacking types : geometric, non-geometric, and lossy JPEG compression.
Scaling image to double size and rescaling to original size modifies intensity slightly because interpolation process only consider 4 pixels in neighbor, therefore the pixel is affected by 4 pixels in surrounding.
Embedding is implemented on LH and HL sub-band (middle frequency sub-bands including a part of low bandwidth), while Gaussian filter is low-pass filter, that means, this signal process will suppress high frequency and keep low frequency. Actually, the quality after using Gaussian filter is decreased, however, it is not enough.
Cropping 25%, that mean, 25% content is removed end replace by 0-bits (black area). We can not extract correctly watermark bit in this area.
When rotate the image, the alteration is increased from the center to border of image, and it affects to all pixels in and image and another reason is the operation of wavelet transform on horizontal and vertical dimension, while the effect of rotation is diagonal.
With histogram equalization, some results is low depend on the contrast of images. Some image have the low contrast (small range in histogram) will be strongly affected by this attacks.
The influence of Gaussian noise on the smooth image (less detail) is less than the image having more details (considering splash sample)
Average filter is the opposite case of Gaussian noise when images which have more detail will be strongly modified by this attack (like as mandrill sample).
Color images, same watermark payload
Comparing with method of Niu at the same conditions, the proposed method is outperform in most of attacking types, except Rotation process.
1. Niu: Contourlet transform. Our method: Wavelet transform.
2. Niu: Embedding on Green channel. Our method: optimal channel selection (3 channels)
3. Niu: Embedding on Low frequencies component. Our method: middle frequencies -> better in imperceptibility
4. Niu: SVM in extraction. Out method: thresholding -> more computation time in training and classification in extraction with Niu method.