Deep Facial Expression Recognition of facial variations using fusion of feature extraction with classification in end to end model

Abstract

Expression recognition is an important direction for computers to understand human emotions and an important aspect of human-computer interaction. Expression recognition refers to the selection of an expression state from a still photo or video sequence to determine the emotional and psychological changes to the character.Spectral Supervised Canonical Correlation Analysis has been used for Feature extraction. For proper classification VGG119 and softmax has been used. Facial variations such as redundant information in image, illumination variance and overfitting have been addressed in this paper. The images have been preprocessed using face detection, data augmentation and image normalization. After down-sampling, Spectral Supervised Canonical Correlation Analysis (SSCCA) holds the dimensions with factor data which constructs affinity matrix that incorporates both the class information and local structure of the data points. Features with having massive discriminative details have been taken. In order to attain low frequency coefficients more effectively the local structural information will be effectively utilized using SSCCA. Data is further provided to VGG19 for proper training. Meanwhile, the proposed method is more effective and robust comparing other methods in the area.