Facial expression recognition (FER) plays a central role in improving human-computer interaction, but its effectiveness is often compromised by variations in facial appearance caused by factors such as head pose, lighting and occlusion. This study investigates the impact of facial alignment and data augmentation on improving the accuracy of FER. We implemented three preprocessing techniques: face alignment using a 68-point landmark detection model, data augmentation through scaling, rotation, and addition of Gaussian noise, and a combined approach of alignment followed by augmentation. Our experiments used the CK+ dataset and selected images from the EXPW dataset to evaluate model performance in different environments. The results showed that while all models achieved comparable accuracies around 85% on the EXPW dataset, the aligned and augmented model did not significantly outperform the others. In particular, the model's performance in recognising sad expressions improved after augmentation, although facial alignment showed a negligible effect, possibly due to the loss of essential features during the alignment process or its limited advantage in complex environments. Conversely, in the simpler CK+ dataset, all models showed reduced accuracy, particularly in distinguishing between sad and angry expressions. To address the observed limitations, we propose to refine the face matching technique by incorporating deep learning methods, and consider a partial matching approach to mitigate overfitting. Future work will focus on enhancing model training by integrating diverse datasets and improving feature selection mechanisms for critical facial features, which are essential for accurate emotion recognition.
Li et al. (Thu,) studied this question.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: