A Facial Expression Recongnition Method Based on Dlib, RI-LBP and ResNet

Abstract

A method of facial expression recognition using a composite feature is proposed. The method combines the expanded Dlib facial feature detector, the rotation-invariant local binary pattern (RI-LBP) and the 50-layer ResNet neural network model (ResNet_50). First, the expanded Dlib was used to locate 83 feature points on the face, obtainting the Dlib feature after preprocessing and dimentionality reduction (PCA). Then, the rotation-invariant LBP feature was extracted from 8 important regions after tilt correction. Furthermore, a 50-layer ResNet neural network was used to extract the low level features from the images. Finally, the three features were combined and extreme learning machine (ELM) was used to classify the composite facial features. The experimental results on Jaffe and CK+ datasets showed that the proposed method performs better compared with other methods.