Abstract
The shape of a hand contains important information regarding the identity for a person. Hand based identification using high-order Zernike moments is a robust and powerful method. But the computation of high-order Zernike moments is very time-consuming. On the other hand, the number of high-order Zernike moments increases quadratically with order causing storage problem; all of them are not relevant and involve redundancy. To overcome this issue, the solution is to select the most discriminative features that are relevant and not redundant. There exists a lot of feature selection algorithms, different algorithms give good performance for different applications, and to choose the one that is effective for this problem is a matter of investigation. We examined a large number of state-of-the-art feature selection methods and found Fast Correlation-Based Filter (FCBF) and Sparse Bayesian Multinomial Logistic Regression (SBMLR) to be the best methods that are efficient and effective in reducing the dimension of the feature space significantly (by 62 %), i.e. the storage requirements and also slightly enhanced recognition rate (from 99.16 +/- 0.44 to 99.42 +/- 0.36).