Occluded Face Recognition: Contrast Correction & Edge Preserving Enhancement based Optimum Features on CelebA Dataset

Abstract

The challenging task of human face identification under variety of occlusions has been interesting and concentrated topic today. The ability of most of the recognition systems drops drastically due to large occluded region and less unoccluded areas. We present an efficient pre-processing and optimum features based occluded face classification system. The occlusions included in the dataset are pose, illumination, age, expressions, hair over face and extend to 40 such difficulties. The pre-processing stage involves two parallel processes and includes contrast measurement-correction and edge-preserving enhancement for the occluded face images. Optimum fine textural features are extracted using Gabor coefficients, Linear Binary Patterns based on Haar Wavelet components and Histogram of Gaussian features. Statistical global features based on first order, wavelet components and color histograms forms the other set of features to represent the whole occluded face region. The work considers 100 celebrities from the CelebA dataset with pose and other occlusions to validate using support vector machine. Experiments analysis carried using the proposed efficient pre-processing and optimum features based face identification system showed improved classification accuracy over other well-known techniques.