Fast Facial Expression Recognition System: Selection of Models

Facial Expression Recognition (FER) is rapidly developing field of Computer Vision and Pattern Recognition directions. FER can be helpful for various purposes: in security systems for aggression recognition, in education for students interests recognition, in marketing for customer satisfaction and in the many other fields. Usually we can distinguish seven common facial expressions for all persons. However, it is often important to know: whether a person is positive or negative. This paper describes the recognition system for facial expression in real time, which defines relatively fast and accurate the positive or negative emotion of the faces in the camera view and selection of the architecture of Deep CNN. This effect is a result of the combination of facial detection algorithms and classification algorithms based on the convolutional neural networks. We have compared different datasets (FER2013 and AffectNet) and provided the experiment results in different cases of the FER models and classes, convolutional layers, and filters. We have found that for 8 classes of FER expressions the architecture model M2 is the best model. It has the best accuracy and works about 2 times faster on GPU and 3 times faster on CPU than M1 model. It has been also found out that training on AffectNet dataset is significantly better than training on FER2013 dataset due to the differences in number of samples in the given datasets