Image interpolation: the original image is downsampled, and then the interpolation method is used to generate the same carrier image as the original image [36\37\38]
Purpose: copyright protection
Disadvantages: zero watermarking (calculation of polar harmonic moments) [Personally, I believe that zero watermarking requires the third party to participate in the process, and usability is not high]< /p>
Current schemes for medical images:
reversible watermarking [4- 6] reversible watermarking
distortion-free watermarking [7,8] distortion-free watermarking
and zero-watermarking [9] zero watermarking
[Jump to "Wavelet Domain Distortion-free Spread Spectrum Robust Watermarking Design"]
JPEG Unknown Cropping ~ okay
Rotation Scaling Gossamer Filtering Good
Median Filtering Gossamer Noise Peppercorn Noise Less Good
Disadvantages
Steps:
Slanting.
Slantle transform - > mean hidden watermark bit - > histogram correction Prevents top and bottom overflow
Former: histogram shifting (attaching min and max point values)
Avoiding top and bottom overflow:
1) Positional map mapping
2) mod 256
3) Empty domain block classification and error correction coding
4) Histogram adjustment (performed in the null domain)
Compare to integer wavelet transform proposed by others
The image as a whole is subjected to the wavelet transform - > carrier subbands Delineate into non-overlapping blocks - > Calculate the maximum of the absolute value of the mean, nmax,T>nmax
In this paper:
SLT transform Error within acceptable limits
Solve top and bottom overflows ->histogram shift (but requires auxiliary information)
bit plane Bit plane operation
Use regional filtering to find blocks with lower variance
Prevent top and bottom overflows by embedding the wrong bit and then correcting it using ECC
To enhance robustness: use Bit Plane + Repeat Embedding + Region Filtering,choose the region with lower variance for embedding
Disadvantages: it can only embed the hash value of a certain hospital logo, i.e., 160 bits, the rest is tampering information and location information, more biased towards tampering with watermarks
For overflowing, it is still recording the location information.
Advantage: recursive dithering
For overflow T, do all-0 and all-1 embedding to get maximum distortion
PSNR = 41
SSIM = 0.96
(Invisible means invisible to the naked eye)
Steps:
RGB to YCBCR (embedding invisible watermarks into the luminance information in the YCbCr color model, the watermark is robust to JPEG lossy compression and other common signal processing operations. At the same time, the use of chrominance information increases the capacity of RDH in the encrypted domain.)
Encrypt each layer of the image
Generate histograms with Cb Cr separately
(nothing is said about how to ensure that the transform is integer)
Embed into luminance information Y
Encrypt Y again with the second person's key
Personally: the experimental data on robustness is fudged
On how to embed into the magnitude of Fourier transform: refer to Authentication of ownership of mixed-domain color images based on efficient robust watermarking (out of 13)
Embedding the same watermark with a mixed-domain
First: Luminance information channel DFT
Second: Chrominance blue differential channel,improved spreading method, embedding the watermark into the contour transform domain.
[Steps]
Luminance component 2D-DFT transform,can get the magnitude and phase
But! DFT is weakly robust to high filament noise etc., CT is better
Choose two radii and you get the area of the ring between them
Watermark 1/0 becomes +/-g
Additive embedding in the amplitude
CT: Embedding in the blue channel (because the color vision of the human eye is less sensitive to blue than it is to red and green)
Kind of hard, didn't read it carefully
Fuck isn't that what I'm going to write about
[Embedding process]
Host image 3D-IWT
LH3 sub-band DCT
Watermark flag also DCT transformed
Watermark DCT transformed result MD5 encrypted
Additionally embedded in the DCT coefficients of the DCT coefficients of the image
Embedding 80-size signature report (BCH encoding) into HL3 [mentioned in 8 articles]
Reverse transformation
Disadvantage: non-blind
Above reference's [10]
Define an offset, where each coefficient is equivalent to being assigned 0/1
Quantization coefficients? is also defined as 2^ l
Process
1) 4D-haar DWT
2) The idea is similar to your own, the coefficients themselves represent the 0/1 information, and only need to be embedded if they are different
3) The embedding is +-? , making the absolute value of the coefficients a bit smaller
Extraction
Inverse process
A thought DWT if the change to the coefficients is +-2^ l, the result after the inverse transformation is also an integer
(there are still such watery articles in the year 9012)
The host image is iso-orthogonal to the random number, which is equivalent to encrypting the ?
LSB embedding
(the only advantage: send the key by email, sort of innovation)
Problems solved:
Aiming at the problem that the feature region selection is not enough to reflect the important information of the image, which leads to weakening of the robustness, a strong robust watermarking algorithm of the scale-space feature region is proposed
A single transform: the resistance is weak. Attack is relatively weak
In this paper, we use to change the scale space of Harris corner points in Harris-Laplace algorithm to obtain the feature points, determine the size of the feature region according to the size of the watermarked image, and select the feature region near the center of gravity of the image and not overlapping each other in the carrier image, and synthesize the feature region matrix. DWT+DCT+SVD in sequence to complete the embedding
Scale space feature point detection
Steps
Advantages: extraction of the feature points of the image, i.e., the embedding position is selected
Disadvantages: need to record the position of the feature region, and it is an additive embedding, which requires the original watermark in order to extract the information
The whole image, B-channel, DWT, HL sub-band. ,HL subbands Divided into 8*8
Each block Fast fwht transform
Result for SVD
Extraction:
Disadvantage: irreversible
Use DCT+SVD
Key determines embedding position
HVS is least sensitive to the blue channel
However, there is also no consider the case where the pixel values are not integer after the inverse transformation. Still hydrology
Contains a characterization of the pixel values of the medical image
Two embeddings: wavelet histogram shift/low distortion overflow processing algorithm (to deal with the overflow problem in the previous step)
Circumferential self-homomorphic mapping: similar to Caesar's cipher
Low-distortion reversible watermarking algorithm: use three predictions from around to get the prediction error p, p+b
It is reversible
Use cyclic redundancy code to determine whether it is tampered with
Low distortion overflow processing: find the overflowed pixel points, process them and generate tampered records, encode the records and embed them in the host image
Overflowed becomes 255, 0
Record the changed values
The content of China's Remainder Theorem<
Steps
Steps
Host image 8*8 chunks
DCT
Randomly select DCT coefficients to embed the watermark bit
CRT the values of the coefficients in the position (two),get p,q
Get D,b,d
To embed 1,d>=8 /D
If this is not satisfied, a correction is needed
Calculate the complexity of each block, and embed complex blocks
(supposedly solves: the problem of top and bottom overflow)
Performed in the spatial domain, in two phases, horizontally and vertically
Embedding method:
Horizontally: increase the value of the pixels of the even rows, and decrease the value of the odd Vertical: decrease the pixel value of even rows, add one to the pixel value of odd rows
Using histogram
Using histogram shrinkage technique to prevent top and bottom overflow
[Previous method]
Lossless compression: drawbacks Compression ratio is low
DE: Requires positional maps
Embedding 0: If the original is an odd, then -1 becomes even
Original even
Graduation! Farewell