Professional Documents
Culture Documents
698
O-7803-6297-7/OO/$10.0002000 IEEE
Authorized licensed use limited to: Ashesi University. Downloaded on December 02,2021 at 18:27:52 UTC from IEEE Xplore. Restrictions apply.
250
200
150
250
200
300
699
Authorized licensed use limited to: Ashesi University. Downloaded on December 02,2021 at 18:27:52 UTC from IEEE Xplore. Restrictions apply.
Figure 4: JPEG compression: quality factor 50 - p = Figure 6: Rotation 30" - p = 0.96 (left); rotation 45"
0.95 (left); quality factor 25 - p = 0.90 (right).; -p = 0.95 (right).
,am, , , , , ,
Figure 5: Filtering: median filter 3 x 3 - p = 0.98 (left); Figure 7: Scaling 50% - p = 0.94 (left); scaling 150%
gaussian filter 3 x 3 - p = 0.95 (right). - p = 0.97 (right).
notch positions and widths, the detection does not need 3. ROBUSTNESS
the original marked image. However, the detection is
more robust if the marked histogram of the intensity The robustness of the proposed watermarking has been
component is available - this usually means storing 256 tested against basic StirMark attacks [9, lo]. In the se-
integers. quel, results obtained on the test image (Fig. 1) are
presented in Fig. 4-10. Histograms of both marked
(dashed line) and attacked image (solid line) are d i s
2.4. Joint marking and enhancement
played. Their similarity is measured by means of their
Marking with notches was chosen in order to preserve correlation coefficient, p.
the visual appearance of the images. Even thus, for The proposed method is robust against JPEG com-
large notch widths a certain increase in image contrast pression for quality factors 2 25. Thus, one obtains
may appear. p = 0.97 for compression at quality factor 90, p = 0.95
Conversely, if image enhancement is of interest, we for quality factor 50 and p = 0.90 for 25 (Fig. 4).
suggest the use of bi-valued watermarks with zero cor- As expected, the method is very robust against me-
responding to notches and a constant value over the dian filtering and robust against linear filtering. The
remaining graylevel scale. For such watermarks, the results obtained with 3 x 3 window filters are presented
marking procedure can be approached as histogram in Fig. 5 - for median, p = 0.98 and for Gaussian
equalization (the specification of a perfectly flat his- filtering, p = 0.95.
togram) followed by watermarking by notches. Obvi- Very good results are obtained as well against geo-
ously, color image enhancement and contrast increase metrical distortions. Thus, for rotation with 30" one
appear due t o histogram equalization on the intensity gets p = 0.96, for 45", p = 0.95 (Fig. 6); for scal-
component. ing at 50% p is 0.94 and at 150% p is 0.97 (Fig. 7).
Two examples are provided in Fig. 3, where, on For centered cropping (75%) one gets p = 0.95 (Fig.
perfectly flat histogram 2 notches (15 graylevels width) 8). Similar results are obtained for rotation scaling, +
have been inserted: centered on levels 105, 175 (left) random removal of rows and columns, translation, etc.
and 155, 225 (right). The 2 marked copies appears vi- The result for color quantization is p = 0.95 (Fig.
sually identical. Using such watermarks provides more 9). We shall mention the results obtained for special-
robustness against attacks. Moreover, no storage of the ized StirMark attacks: randomization and bending -
original marked histogram is needed in the detection p = 0.95, FMLR - p = 0.94 (Fig. 10). Lower robust-
stage. ness was obtained against sharpening ( p = 0.89).
700
Authorized licensed use limited to: Ashesi University. Downloaded on December 02,2021 at 18:27:52 UTC from IEEE Xplore. Restrictions apply.
Figure 8: Centered cropping 70% - p = 0.95. Figure 10: StirMark with randomization and bending
- p = 0.95 (left); FMLR - p = 0.94 (right);
70 I
Authorized licensed use limited to: Ashesi University. Downloaded on December 02,2021 at 18:27:52 UTC from IEEE Xplore. Restrictions apply.