Log-domain Polynomial Filters for Illumination-Robust Face Recognition

Yinyan Jiang,Yong Wu,Weifeng Li,Longbiao Wang,Qingmin Liao

DOI: https://doi.org/10.1109/icassp.2014.6853647

2014-01-01

Abstract:This paper proposes a novel face image descriptor local surface pattern (LSP) for illumination-robust face recognition. It is assumed that the discrete array of pixel values comes about by sampling an underlying smooth surface on the domain of the image. The proposed method efficiently estimates the underlying local surface information, which is approximately represented as linear projection coefficients of the pixels in a local patch. Thus, by filtering local image patches using the polynomial filters and binarizing the filter responses via thresholding, the method can compute a binary code for each pixel in the face image. Then the distribution of the code over suitable image regions is used for face representation. Furthermore, we prove that applying zero-mean filters in logdomain may enable the responses to be more robust to illumination variations. The experimental results on Extended Yale-B and FERET fc databases illustrate the effectiveness of our proposed method in illumination-robust face recognition.