Cryptography Reference
In-Depth Information
the feature-annotation-copyright-based query scheme, we can only search the
desired images having the same copyright and the same annotation.
11.4.4 Experimental Results
We did the experiments on a test database of 1000 images in 10 classes, each
class including 100 images. They are with size of 384256 or 256384. Dur-
ing the o
ine process, we extract and embed three kinds of features. These
are, one Global Invariant Feature, four Hu Moments and 12 Histogram-Based
Moments (this is for the three R, G, and B Color Components) for each
database image. For the Global Invariant Feature, we use n = 10000 and
f (M)=M (0, 1)M (2, 0). We embed a 4848 sized binary copyright wa-
termark, 17 floating-typed (68 bytes) feature values and 16-bytes annotation
text. In our system, we can embed up to 288 bytes. In the experiment, we
view the annotation text as part of the Feature Watermark. We combine the
feature watermark and the Copyright Watermark to construct the watermark
to be embedded for each database image. We then embed the watermark in
the corresponding image. After we obtain a watermarked image database, we
can perform the online retrieval to obtain answers to various queries.
For a query based on features, we show an example of retrieval results
in Fig. 11.22. The average precision and recall for each class is shown in
Table 11.2. The average precision and recall for each class can be obtained
as follows. We first randomly select ten images from the class. We then use
each image as the query image. For each query image, we obtain the recall
by finding the ratio of returned images in this class in the first 100 returned
images. We then find the position of the first returned image which is not
in this class and divide it by 100 to obtain the precision. After finding ten
recalls and ten precision values, we average them to get the average recall and
precision.
For the image retrieval system, the most important operation to which our
system should be able to resist is the high-quality compression. Other attacks
are not so important. To show the performance of the system in resisting the
JPEG compression, we give some results in Figs. 11.23 and 11.24. Under the
condition that we can extract the watermark which is more than 80% similar
to the original embedded information, Fig. 11.23 gives the minimum com-
pression quality factors to which the system can resist with. It uses a number
of different modulation steps. Fig.11.24 shows the average watermarked im-
age qualities obtained under different modulation steps. In previous sections,
we have mentioned that the extracted feature watermark should not be re-
sponsible for bit loss. With regard to this, in the case of 100% recovery, the
experimental results show that the average lowest JPEG compression quality
factor to which the feature watermark can resist is 90.
