Facial Expression Recognition using Transfer Learning and Fine-tuning Strategies: A Comparative Study

Facial expression recognition goes from laboratory-controlled to challenging in the wild due to various unconstrained situations. Due to human-computer interaction technology advancements, people pay increasing attention to facial expression recognition techniques. In this paper, we saw some popular convolution networks architectures and surveyed their development to FER. These architectures are MobileNet, VGG16, VGG19, DenseNet121, Inception-v3, and Xception. We also looked at various convolution neural network theories. We implemented our work on the Facial Expression Recognition 2013 dataset, one of the most important datasets of human faces. We illustrated the power of employing transfer learning and fine-tuning for training each selected model. The evaluations clarified that the fine-tuning outperforms the transfer learning and the VGG16 model has the highest accuracy of 0.6446.