Convolutional Neural Networks-based Evaluation for the FER-2013 and Revised CK+ Datasets

This study seeks to enhance the efficiency of facial Emotion recognition, addressing challenges posed by variations in facial morphologies, lighting conditions, and occlusions, which are critical for improving human-machine interaction. Convolutional Neural Networks (CNN) were utilized to train and evaluate models on the FER-2013 and revised CK+ datasets, incorporating data preprocessing techniques like normalization and data augmentation, and varying training parameters such as epochs and batch sizes. The revised CK+ dataset, trained for 50 iterations using a batch size of 512, showed superior performance with the lowest training loss (0.0042) and testing loss (0.0222), as well as the highest training accuracy (0.9982) and testing accuracy (0.9947). In contrast, the FER-2013 dataset with 20 epochs exhibited the highest training loss (0.1942) and testing loss (0.7067), along with the lowest training accuracy (0.9319) and testing accuracy (0.7934). We need to increase the epochs for the FER-2013 dataset to 50 significantly improve its performance, highlighting the importance of optimal training parameters and dataset refinement. The findings indicate that a well-designed CNN model can greatly improve facial expression recognition accuracy, with the revised CK+ dataset proving especially effective. Future research will focus on further refining datasets, exploring more complex datasets, and implementing real-time applications to ensure model robustness and generalizability.