A Combined Local Deep-Feature Alignment and Analytic Gabor Feed forward Network for Face Recognition

Abstract

This paper combined an unsupervised framework called as Local Deep-Feature Alignment (LDFA) for dimension reduction and Analytic Gabor Feedforward Network (AGFN) for efficient face recognition. An affine transformation is exploited for aligning the local deep features of each neighborhood with the global features. A new data sample is mapped into the low-dimensional subspace. The Stacked Contractive Auto-Encoder (SCAE) maintains the locality characteristics in the neighborhood. Regularization term of each SCAE facilitates estimation of parameters from a neighborhood that does not contain more data. The difficulty in robust feature learning is reduced by minimizing the variations of the local embedding function due to the data similarity among each neighborhood. As the local features are learned from minimum data in a neighborhood, the LDFA method yields better performance. The AGFN works directly on the input raw face images and generates Gabor features at the hidden layer. Several sets of features obtained at different orientations and scales are combined. The proposed method achieves high recognition accuracy, true positive, accuracy and minimum execution time and false positive than the existing face recognition techniques.

 Index Terms—Analytic Gabor Feedforward Network (AGFN), Face Detection and Recognition, Local Deep-Feature Alignment (LDFA), SVM Classifier