Extract the Pain Features from Facial Expressions by Utilizing Two DCNN Model

Adopting facial expressions in automatic pain recognition is a complicated matter that has intrigued the interest of the scientific community. This paper provides a performance comparison between different Deep Convolution Neural Network (DCNN) architectures with different deepnesses (number of layers and connections): ResNet-50 and DenseNet-121. These models were using as feature extractors by removing the fully connected layer and freezing its convolution layers. Then the features vector is fed to four types of classification algorithms: SVM (Support Vector Machine), KNN (k Nearest Neighbor), LR (Logistic Regression), and building a new DNN (Deep Neural Network) classifier from scratch. The data set is collected from 110 subjects adults and teenagers (71 males and 39 females) based on natural environment and different times. The best classification results is 96.15% by utilizing the DenseNet-121 model in mirage it with different classifiers. The most important things that must be perceived from this result are the depth of the network and the proper choice of transfer learning technique.