[1] [2] [3] [4] [5] [6] [7] [8] [9] [10] [11] [12] [13] [14] [15] [16]

Journal of Information Science and Engineering, Vol. 37 No. 1, pp. 15-39

A Novel Dual CNN Architecture with LogicMax for Facial Expression Recognition

Department of Electronics and Communication
National Institute of Technology, Warangal
Telangana, 506004 India
E-mail: ahadit.ab.ml@gmail.com; ravikumar@nitw.ac.in

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 F1 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.

Keywords: convolutional neural networks, transfer learning, facial action coding system, action units, Pearson correlation, data augmentation, dlib facial landmark predictor, vgg16, logicMax

  Retrieve PDF document (JISE_202101_02.pdf)