Recognition Oriented Feature Hallucination for Low Resolution Face Images

Abstract

In face recognition, Low Resolution (LR) images will lead to the decline of the recognition rate. In this paper, we propose a novel recognition oriented feature hallucination method to map the features of a LR facial image to its High Resolution (HR) version. We extract the principal component analysis (PCA) features of LR and HR face images. Then, canonical correlation analysis is applied to establish the coherent subspaces between the PCA features of the LR and HR face images. Furthermore, a recognition rate guided prediction model is proposed to map the LR features to the HR version, which is employed an adaptive Piecewise Kernel Partial Least Squares (P-KPLS) predictor. Finally, a weighted combination of the hallucinated PCA features and the Local Binary Pattern Histogram (LBPH) features are adopted for face recognition. Experimental results show that the proposed method has a superior recognition rate.