Performance Comparison of Transfer Learning and Training from Scratch Approaches for Deep Facial Expression Recognition

Abstract

Convolutional neural networks (CNN) are often used in many areas such as object detection/recognition, biomedical image analysis, driver monitoring, facial expression recognition, etc. Owing to its popularity and performance, many networks and approaches have been developed by the scientists. Transfer learning is one of these approaches. With this approach, many machine learning problems can be solved fast and successfully. In this study, alexnet and vgg16 networks are used to observe the effectiveness of transfer learning and training from scratch methods on the facial expression recognition task. To perform this task, four scenarios are built. In every scenario; training set, validation set, testing set, and parameter numbers of the networks are same. Just learning approaches (transfer learning, training from scratch) and network types (alexnet, vgg16) are different. Experimental results show that the transfer learning approach achieves higher success rates and shorter successful training time compared to the training from scratch approach. The best average accuracy is obtained from transfer learning with vgg16 network on RaFD database. Disgust, fear and happy expressions are classified impeccably, and general classification accuracy is %98.33 for vgg16 with transfer learning.