Improving the Recognition Performance of Moment Features by Selection - Feature Selection for Data and Pattern Recognition - page 316

Table 13.5 Recognition performance of moment features subsets for the Radboud dataset
Radboud dataset
Number of moments Moment type Recognition rate (%) Selection method

5 Zernike 48.21 GA

10 Zernike 55.75 GA

15 Zernike 61.96 GA

20 Zernike 61.38 GA

25 Zernike 62.14 GA

30 Zernike 54.10 GA

40 Zernike 53.30 Relief

50 Zernike 53.66 Relief

60 Zernike 53.30 Relief

70 Zernike 51.47 No Sel./Relief

Table 13.6 Best configuration for each dataset
Dataset Number of moments Moment type Recognition rate (%) Selection method

Ya l e 15 Zernike 88.00 GA

Terravic 5 All 100.00 No Sel./GA

JAFFE 10 Krawtchouk 79.90 GA

Radboud 25 Zernike 62.14 GA

accuracy. However, Krawtchouk moments show a significant performance leading
to the conclusion that the locality property plays an important role to capture the
local characteristics of the patterns. More work has to be done in this direction of
describing the local information by the method of moments.

13.6 Conclusions

A detailed discussion of the main properties of the most representative image orthog-
onal moment families was presented in the previous sections. Through an in depth
analysis of the representation capabilities of the orthogonal moments, the need for
selection of moment features for improved recognition accuracy is highlighted.
Finally, an extensive experimental study on well known benchmark datasets has
resulted in useful conclusions regarding the initial assertion of moment's selec-
tion and the description capability of each moment family. The GA-based selection
method has shown superior performance to the Relief algorithm, mainly for low
number of moments, while for high number of features the latter algorithm seems to
be the suitable choice.