Page 53 - 2024-Vol20-Issue2
P. 53
49 | Alobaidi & Mikhael
The energy residual, F(a, ß ), the distinction lies in minimiz- manner:
ing the cost function, which is calculated as the discrepancy
between the initial energy and the energies preserved within s(a, b) = sab +C3)
each domain. In particular, F(a, ß ) is calculated as follows: sasb +C3
F(a, ß ) = [C1]2 - [T2,1(C2)]2 - [T3,1(C3)]2 c(a, b) = 2sasb +C2
sa2 + sb2 +C2
(10)
C2 = (K2 * L)2 (16)
where [ ]2 is the element-wise square. The process begins by l(a, b) = 2µaµb +C1
utilizing a Steepest Descent Algorithm [22] to decrease the µa2 + µb2 +C1
remaining error. Once the iteration concludes, a designated
count of coefficients is preserved in two distinct domains: the C1 = (K1 * L)2
Cosine and the Haar domains. The resulting feature vector
for each signal (here, it is the cover image) is obtained by where s depicts standard deviation, µ represents mean, s 2
combining these retained coefficients together. represents variance, K1 = 0.01, L equals to one, C3 is a small
The parameters for the Training phase are as follows. The constant, and K2 = 0.03.
weight matrices a is populated with 0.5 while ß is initialized
with 0.3. The updating equations in every iteration are as 2) Root Mean Squared Error (RMSE)
follows [23]: Mean Squared Error, one of the Regression metrics, measures
the average squared difference between the pixel values of
ax,y(n + 1) = ax,y(n) - µax,y ?ax,y F (11) the original signal and the reconstructed/compressed signal.
ß x, y(n + 1) = ß x, y(n) - µßx,y?ßx,yF (12) It provides a quantitative measure of the overall distortion
between the two signals. The Root MSE (RMSE), the square
where x, and y span the entire domain and depending on root of the MSE, is calculated as follows:
ax,y and ßx,y are elements in [a] and [ß ] respectively, n is the
iteration index, and µ is the converging factor. The converging RMSE = m 1 n ? ?(I(x, y) - K (x, y))2 (17)
factors, µax,y and µßx,y, are calculated in the following fashion: ×
m n
µa = F(n) (13) where I(x, y) represents the pixel value of the original signal
(14) at position (x, y), K(x, y) represents the pixel value of the
N-1 N-1 [?ax,y F]2 reconstructed/compressed signal at the same position, and (m
* n) is the total number of pixels in the image. A lower RMSE
?x=0 ?y=0 value indicates a smaller average difference and, therefore,
better reconstruction or compression quality. However, RMSE
µß = F(n) alone may not provide a perceptually meaningful measure of
?xN=-01 ?Ny=-01[?ßx,y F]2 quality, as it does not consider the human visual system’s
sensitivity to different image characteristics.
E. Performance Metrics
Performance metrics [24] are measurements used to evaluate 3) Peak Signal-to-Noise Ratio (PSNR)
the effectiveness, efficiency, accuracy, or quality of a system, PSNR is a logarithmic measure that relates the maximum
process, algorithm, or model. The choice of performance possible power of a signal (in this case, the maximum pos-
metrics depends on the specific task or application. sible pixel value) to the power of the noise (the difference
between the original and reconstructed/compressed signals).
1) Structural Similarity Index (SSIM) It is expressed in decibels (dB). The formula for PSNR is:
To measure similarity between parts of same or different im-
ages, SSIM is employed. Post-processing quantitative judg- PSNR = 10 * log10(MAX2/MSE) (18)
ment of the change of the structure of these parts is measured
by SSIM. Structure, contrast, and luminance are the segment
ofSSIM. SSIM ? [-1, 1], and the maximm limit is reached
when image parts are identical. The calculation of SSIM is:
SSIM(a, b) = [s(a, b) * c(a, b) * l(a, b)] (15) where MAX is the maximum pixel value (e.g., 255 for an
8-bit grayscale image). PSNR provides a more perceptually
where a and b are input images (or blocks) under comparison, relevant measure of quality because it takes into account the
s(a, b) is structure component, c(a, b) equals contrast, and l(a, dynamic range of the pixel values and is logarithmic. A higher
b) is luminance. These factors are calculated in the following PSNR value indicates better quality, as it indicates a smaller
ratio of noise to the maximum signal power.