Improving the Recognition Performance of Moment Features by Selection - Feature Selection for Data and Pattern Recognition - page 315

Table 13.3 Recognition performance of moment features subsets for the Terravic dataset
Terravic dataset
Number of moments Moment type Recognition rate (%) Selection method

5 All 100.00 NoSel./GA

10 All 100.00 NoSel./GA

15 All 100.00 NoSel./GA

20 All 100.00 NoSel./GA

25 All 100.00 NoSel./GA

30 All 100.00 NoSel./Relief

40 All 100.00 NoSel./Relief

50 All 100.00 NoSel./Relief

60 All 100.00 NoSel./Relief

70 All 100.00 NoSel./Relief

Table 13.4 Recognition performance of moment features subsets for the JAFFE dataset
JAFFE Dataset
Number of moments Moment type Recognition rate (%) Selection method

5 Legendre 71.66 GA

10 Krawtchouk 79.90 GA

15 Legendre 78.90 GA

20 Krawtchouk 77.47 GA

25 Krawtchouk 69.33 GA

30 Krawtchouk 53.92 GA

40 Dual Hahn 47.35 Relief

50 Zernike 46.88 Relief

60 Gaussian-Hermite 46.80 Relief

70 Legendre 46.88 NoSel.

From the above results, it can be observed that the increase of the number of
moments used to discriminate the patterns does not always improve the recognition
accuracy. In almost all the cases a subset of 10-25 moment features is able to achieve
the highest recognition rate.
In order to draw a conclusion regarding the optimal settings, ensuring the best
solution to each dataset, the most effective configuration in each case is summarized
in Table 13.6 .
The results of Table 13.6 showagain the outperformance of theGA-based selection
method over the Relief one, while its recognition accuracy is in agreement with the
state of the art methods [ 11 , 22 , 25 ]. As far as the performance of themoment families
is concerned, it is obvious that Zernike moments are the most efficient family, while a
moments' subset of size lower than 25 is adequate to ensure an acceptable recognition