A Transfer Learning Framework for Facial Emotion Recognition: Leveraging Pre-Trained Convolutional Neural Networks

Emotions represent a fundamental aspect of human existence, emotions encompass a complex interaction of mental, physical, and behavioral reactions to external stimuli. Emotions are adaptive mechanisms found in the fields of affective neuroscience and psychology. There are other categories into which they might be divided, including fear, joy, sorrow, rage, and surprise. In particular, facial expressions are an essential tool for interpreting and expressing various emotional states. Face emotion recognition task classifies seven categories of emotion i.e. happy, sad, disgust, fear, surprise, anger, neutral. This study is to demonstrate the efficacy of transfer learning and the usage of pretrained models to recognize seven distinct types of emotions. The newly defined-connected layers and output layer for classification, as well as the Convolutional and Pooling layers of pre-trained models—ResNet50, DenseNet121, EfficientNetB5, and Xception— are the foundation of the system that is being shown. The proposed framework is specifically created for the Face Emotion Recognition dataset (FER). The EfficientNetB5 model with the biggest impact outperformed the other four pre-trained models in the study, obtaining the highest training accuracy of 86.23% and testing accuracy of 84.39%.