This document summarizes a DCT-based watermarking technique that allows detection of the watermark from a corrupted image without needing the original uncorrupted image. The technique embeds a watermark by modifying mid-frequency DCT coefficients of an image. During detection, the watermark is recovered by correlating the extracted DCT coefficients against different potential watermarks and selecting the one with the highest correlation. This avoids the need for the original image by directly correlating against watermarks rather than trying to estimate the original coefficients.
1. DCT Based watermarking recovering without
restoring to the uncorrupted Original Image
Umair Amjad umairamjadawan@gmail.com
2. Agenda
Watermark embedding
Watermark detection
DCT-basedWatermark Recovering
without Resorting to the Uncorrupted Original Image
A. Piva, M. Barni, F. Bartolini, V. Cappellini
Dipartimento di Ingegneria Elettronica
Universit`a di Firenze
via S. Marta 3, 50139, Firenze, Italy
e-mail piva@cosimo.die.unifi.it
3. Watermark Embedding
Read image. Lets say, I.
DCT transform is applied.
DCT coefficients are reordered into a zigzag scan.
It is now impossible for the decoder to determine the position of the coefficients with the
largest magnitude, since non-marked image is no longer available.
T = {t1, t2, . . . . . . , tL, tL+1, . . . . . . , tL+M}
To get perceptual invisibility we skipped first L coefficients.
Now, watermark X = {x1, x2, . . . . . . , xM} (choose among pseudo-random sequences) is
embedded in the last M numbers, to obtain a new vector.
T’ = {t1, t2, . . . . . . , tL, t’L+1, . . . . . . , t’L+M}
According to the rule:
t’L+i = tL+i + α |tL+i| xi
Where i = 1, . . . . . , M
Now the vector T’ is inserted in the inverse zigzag scan and the inverse DCT algorithm is
performed, obtaining the watermarked image I’.
4. Watermark Detection
Read watermarked image. Lets say I*.
DCT transform is applied.
Perform zigzag scan.
Now coefficient (L+1)th to (L+M)th are selected to generate a vector.
T* = {t*L+1, t*L+2,. . . . . . , t*L+M}
The correlation between the corrupted coefficients T*, and the mark itself is taken as a
measure of the mark presence.
For this we generate 1000 random fake watermarks.
Correlation is computed for each of the marks and that with the largest correlation is
assumed to be the one really present in the image.