I am getting better accuracy with Euclidean distance for Zernike features than SVM or ANFIS. Is it possible?

I have extracted Zernike Moment features from ORL face database and classified using SVM (libsvm tool), ANFIS of MATLAB, Euclidean distance and Backpropagation Algorithm. I am getting 93.5%, 90.5%, 96% and 90% accuracy respectively for each classifier. Highest accuracy has been obtained using simplest Euclidean distance measure. Is it really possible or I may be doing something wrong? Because as I see literature than SVM is definitely a better classifier. But in my case this is not happening. Please any suggestions/help?

Answer 1

[Walter Roberson]

Zernike Moment features are rotational invariant (at least if you take their magnitude). Euclidean distance is rotational invariant. SVM involves finding a hyperplane in some space to discriminate between features, so SVM is directional in that hyperplane; whether it is rotational invariant in the original space depends on the mapping between the original space and the space the hyperplane is in... but chances are that it is not rotational invariant.

http://murphylab.web.cmu.edu/publications/boland/boland_node33.html

[Geetika]

Thank you very much for your answer. So we can say that ZM features might give better results with Euclidean distance over SVM. Can you give some comments on such behavior with ANFIS?