Occluded face recognition using optimum features based on efficient preprocessing and machine learning

Abstract

Face occlusion is a major challenge for today’s face recognition system. The occlusions included in current datasets are pose, illumination, age, expressions, and natural and artificial obstacles over the face and extend to 40 such difficulties. The performance of even a robust system would fail when the occluded area is large as compared to the un-occluded region. We extracted traditional features based on an efficient preprocessing mechanism and classified them using machine learning over scarce datasets. The preprocessing stage involves obtaining two sets of images based on contrast correction and anisotropic filtering and then averaging them. Optimum quality features are then extracted from the mean color and grayscale image using diverse descriptors such as Gabor, Linear Binary Patterns based on Haar Wavelet components, Histogram of Gaussian features, Statistical global features based on first order, wavelet components, and color histograms. The proposed work outperforms state of art techniques concerning classification accuracy obtained using Support Vector Machine.