MASTER THESIS Video Watermarking - Computer Graphics Group ...

as a part of the whole process. Then substituting detected watermark element
values by picture sample values, the evolution of the correlation sum is:
where A and B stand for contributions to the sum from the picture and from the
watermark. Let us assume that A is zero because of independence of the
pseudo-random signal and the picture, then:
In practice, A is not exactly zero, thus an error is included. The decision
process in the framework is invoked – when Ci is greater than 0, the value of the
detected hidden information bit is 1, and when Ci is lower than 0, the value is -1.
Ii is either -1 or 1 and a is greater than 0, therefore the sign of Ii sets the sign of
Ci and the detected hidden information bit is determined correctly. The greater
the weight factor a is the greater the tolerance to the error of A is provided.
The probability of detection success may be increased by applying a
high-pass filter to the sequence before the detection process in order to filter out
the host signal and keep the watermark signal alone.
4.3.2 Block Watermark
Block watermarking method belongs to frequency domain techniques. The
method consists in coding one watermark element into one block of a macroblock
residual.
Only 4×4 blocks are supported because of the following. The partitioning of
macroblock residuals into blocks may change when the video sequence undergoes
any video signal processing operation. The simplest example is recompression
with different parameters.
The problem occurs when the watermark element has been embedded into a
macroblock partitioned into 16 4×4 blocks and the partitioning has changed to 4
8×8 blocks, or vice versa. Changes made by watermark embedding in one
partitioning are basically undetectable in the other partitioning because the
transforms are not equivalent in terms of transform coefficient values. It would
be possible to convert the blocks to the former partitioning before detection but it
is not obvious which partitioning is the former one.
Therefore, the conversion to one type of partitioning has to be applied before
embedding. After embedding, the partitioning is converted back to the former
type in order to preserve macroblock properties. In the detection process, the
conversion to the same type as in the embedding process is applied before
31
(10)
(11)