Xception Based Transfer Learning Combined With Residual Block for Facial Emotion Recognition

Abstract

Human emotion plays a crucial role in communication, making its recognition an essential task. In recent years, transfer learning and residual blocks have emerged as effective methods for improving image classification performance. This paper aims to develop an emotion recognition model by leveraging the Xception architecture and incorporating residual blocks from ResNet. To achieve this, we made modifications to both the model architecture and the residual block architecture. Additionally, we utilized two different validation datasets. The second validation dataset involved image augmentation techniques, introducing additional variations and noises. Our study demonstrates the benefits of integrating residual blocks into the Xception-based model, achieving a validation accuracy of 65.78% on the first dataset and 60.56% on the second, surpassing the accuracy of the standard Xception model. However, during experimentation, we observed that all tested models exhibited overfitting, with validation accuracy significantly lower than training accuracy, partly attributed to the limited number of images in the datasets. As for the future works, we intend to address dataset imbalance through image augmentation and implement dropout layers to prevent overfitting.