Optimal feature selection based on mutual information for face recognition

Gabor face representation has been getting popular in face recognition applications. However, it also suffers from the high dimensional data containing diverse redundancy and different random noises. To utilize the Gabor feature for efficient face recognition, a new Gabor feature selection method is proposed. Firstly, the Gabor feature differences between every two face images within a training data set are calculated and grouped into two categories: intra-individual set and extra-individual set. Then the rank of discriminating capabilities of features can be estimated by evaluating the classification error on intra-set and extra-set based on weak classifier built by single feature. The Gabor features with small errors were selected. And at the same time, the mutual information between the candidate feature and the selected features was examined. As a result, the non-effective features carrying information already captured by the selected features will be excluded. The features thus selected are both accurate and non-redundant. Finally, the selected Gabor features were classified by PCA and LDA for final face recognition. The experiments on CAS-PEAL large-scale Chinese face database show that the proposed method can greatly reduce the dimensionality of Gabor features and effectively increase the recognition accuracy.