Face Angle Identification with Ensemble Learning

Abstract

The current face recognition methods may not work well on non-upright faces due to the heterogeneity of the training set. This study proposes an ensemble model to identify the angle of a face. The face is then rotated according to the predicted angle before feeding to face recognition. Our proposed model takes advantages of Divide-and-Conquer methods which breaks down a complicated problem to several simple tasks. A number of based classifier, i.e. Convolutional Neural Network, with a simple structure is used to classify whether a face is in a given range or not. The final angle prediction is determined by the majority voting. The experimental results suggest that our method achieve excellent performance in terms of accuracy and speed in face angle identification.