Face Recognition Revisited on Pose, Alignment, Color, Illumination and Expression-PyTen

Growing interest in intelligent human computer interactions has motivated a recent surge in research on problems such as pose estimation, illumination variation, color differences, alignment distinction and expression variation. Human faces are highly non-rigid objects with high degree of variability in pose, color, expression, alignment angles and illumination conditions and most face recognition algorithms (not all), are designed to work best with well aligned, well illuminated, and frontal pose face images. An optimal face representation should be discriminative, robust, compact and easy to implement. The conventional pipeline of face representation consists of image pre-processing, extraction, alignment, representation and classification. Our approach is based on feature sharing structure of deep network called Pyramid CNN (Pyramid Convolutional Neural Network) which has known to adopt a greedy filter and down sampling approach for a fast and computation efficient training procedure. CNN learns representation of the face utilized by recognition algorithms in later stages. The color values of face images are normalized to RGB color space to reduce the lightning effect in normalization process. We use Field proposed Log Gabor filters for feature extraction which allows more information to be captured in high frequency domains with desirable high-pass characteristics. Using feature sharing Pyramid CNN we are able to achieve competitive accuracy on LFW database