Deep learning based advanced facial expression recognition system

Facial emotion expression recognition is a field of research that comprises the classification of face emotions of humans by expressions on their faces. It can be used in many different applications including intelligent human-computer interaction, biometric security, robotics and depression, and clinical medicine for autism, and mental health problems. This thesis explores and analysis advanced techniques for facial expression recognition (FER) and develops intelligence systems for practical applications. In this study, several deep learning-based frameworks have been developed to improve FER accuracy. Three main types of pre-trained networks (AlexNet, GoogleNet, and SqueezeNet) are utilized for feature extraction purposes at a certain layer. Moreover, k-nearest neighbors (KNN), Support Vector Machine (SVM), and Decision Tree neural networks algorithms are employed as a classifier for each type of pre-trained network. Two datasets are used in this research including a large number of images representing Seven types of facial expressions. The maximum accuracy obtained for GoogleNet with SVM is 91.09, SqueezeNet with SVM 98.2766, and AlexNet with KNN at about 100%. The results obtained indicate that we can use a pre-trained network as feature extraction which provides a pre-training time and resources for best classification results.