Automatic Facial Expression Recognition System Using Deep Gradient Framework

Abstract

In human face, some regions are utilized to discriminate with one another. This region wise facial expression identification is done by many research scholars. Some of local facial descriptors had shown their efficacy in facial expression recognition (FER). In this paper we mainly concentrate on extracting more informative and discriminative features of face using Sobel , Schaar, Speeded-Up Robust Features (SURF) and  Histogram of Oriented Gradient (HOG) descriptors. These error free and more accurate gradient images are given as input for deep learning framework namely ResNet18 which is most sophisticated deep framework for classification. For four gradient feature maps considered as input for four ResNet18 frameworks and trained individually. These four ResNet18 networks responses are fusion using squeeze and excitation method. This deep gradient framework derived more optimal feature set for accurate facial expression recognition.  We consider seven expressions namely, happiness, sadness, surprise, fear, anger, neutral and disgust. For better results used person-independent and leave-one-out concept in training and testing formation. The experiments conducted on different benchmark databases with fusion feature set obtained from ResNet18 illustrate the effectiveness of the proposed methodology.

Keywords- leave-one-out, independent, expressions, region, fusion