Information Technology Reference
In-Depth Information
PSNR = 30 dB. In this case PSWR = 45.46dB, obtained from Figure 3. This was in
good agreement with simulation results.
Three common types of attack were tested in the experiment. They were low-pass
filtering, noise interference, and JPEG compression.
Table 3 gives character-extraction error rates at different noise levels. In the ex-
periment, 50 tests were carried out with a total of 800 embedded characters to produce
the statistical results. The experimental results are in line with the theoretical calcula-
tion based on Equations (21), as shown in Figure 4.
Table 3.
Character-extraction error rate at different noise levels
PSNR (dB)
28
30
32
34
36
Character-extraction error rate (%)
2.62
1.13
0.75
0.50
0.38
5
4.5
4
3.5
3
2.5
2
1.5
1
0.5
0
26
28
30
32
34
36
38
Fig. 4.
Character-extraction error rate: theoretical calculation (solid line) and simulation results
(asterisks)
PSNR caused by noise (dB)
3 Gaussian convolution mask. The strength
of attack was characterized by the standard deviation
Blurring was implemented using a 3
×
. Table 4 shows the simulation
results. Table 5 gives the character-extraction errors after JPEG compression. As
expected, the measured error rates increased with increase of the strength of attacks.
σ
Table 4.
Character-extraction error rate after low-pass filtering
0.7
0.6
0.5
0.4
0.3
Standard deviation of Gaussian mask (σ)
Peak signal-to-distortion ratio (dB)
31.7
33.4
36.9
45.0
65.2
Character-extraction error rate (%)
2.88
1.25
0.75
0.25
0.13
Table 5.
Character-extraction error rate after JPEG compression
Quality factor (Q)
40
50
60
70
80
Compression ratio
9.82
8.60
7.61
6.45
5.16
Character-extraction error rate (%)
3.62
1.87
1.00
0.63
0.37
