Face Recognition Based On Multiscale Information Fusion

Biometric technologies have advanced rapidly in the past twenty years. And many feature descriptors were developed during this period of time, such as HOG, MSER, Gabor and LBP. But, facial recognition is still being questioned about its effectiveness even though these descriptors are used.In order to make these feature descriptors play better and more sufficient roles, it is necessary to introduce feature fusion technologies including fusing the features from different descriptors and fusing the features from the same descriptor. Thus this paper focuses on multiscale feature fusion. The main contents of our work can be summarized as follows:(1) We presented an analysis of multiscale-based approach from feature engineering perspective and proposed a general multiscale information fusion method of utilizing the features from the same descriptor more fully. In this method, more than one feature vectors are extracted from each face image, which correspond to different combinations of parameter values. Then these vectors are weighted and fused into a distance function. Our method have been evaluated with LBP and HOG on AR and FERET face databases. Results show that this method does indeed produce better recognition performance.(2) However, this method introduces some additional parameters that need to be optimized, making it more difficult to use than the original LBP. Thus, a self-adaptive feature fusion method is proposed which introduces a mean-based LBP (mLBP) and extends it by adding a parameters adaptive module. This method involves four steps. Firstly, a large number of initial features are generated. Secondly, a scoring method based on Fisher criterion is used to evaluate the discriminative ability of different groups of features. Thirdly, a novel prism volume model is presented for optimum parameter set selection. Finally a distance function is used to fuse the resulting multiscale features for classification. Experiments on the ORL and AR face databases have shown that the proposed method can learn parameters in a self-adaptive way, and can produce excellent classification performance comparable to the case when the parameters are optimized in a brute force approach.