A Novel Dual CNN Architecture with LogicMax for Facial Expression Recognition

Facial expressions convey important features for recognizing human emotions. It is a challenging task to classify accurate facial expressions due to high intra-class correlation. Conventional methods depend on the classification of handcrafted features like scale-invariant feature transform and local binary patterns to predict the emotion. In recent years, deep learning techniques are used to boost the accuracy of FER models. Although it has improved the accuracy in standard datasets, FER models have to consider problems like face occlusion and intra-class variance. In this paper, we have used two convolutional neural networks which have vgg16 architecture as a base network using transfer learning. This paper explains the method to tackle issues on classifying high intra-class correlated facial expressions through an in-depth investigation of the Facial Action Coding System (FACS) action units. We have used a novel LogicMax layer at the end of the model to boost the accuracy of the FER model. Classification metrics like Accuracy, Precision, Recall, and Fl score are calculated for evaluating the model performance on CK+ and JAFFE datasets. The model is tested using 10-fold cross-validation and the obtained classification accuracy rate of 98.62% and 94.86% on CK+ and JAFFE datasets respectively. The experimental results also include a feature map visualization of 64 convolutional filters of the two convolutional neural networks.

Automated summary

This paper presents a novel method to enhance the accuracy of facial expression recognition models by using transfer learning and a LogicMax layer. Through an in-depth investigation of the Facial Action Coding System (FACS) action units, the paper addresses the issue of classifying highly correlated facial expressions. The experimental results show promising classification accuracy rates on standard datasets CK+ and JAFFE.