FACIAL EMOTION RECOGNITION USING DEEP LEARNING AND MACHINE LEARNING: A COMPARATIVE STUDY ON CK+ AND FER2013 DATASETS

Abstract

Facial Emotion Recognition (FER) is a critical aspect of human-computer interaction, healthcare, and security. Still, the ability to detect the presence of emotion and classify it depends on the strength of feature extraction and precise identification in various circumstances. This paper compares the Support Vector Machines (SVM) and Convolutional Neural Networks (CNN) on two benchmark datasets, CK+ and FER2013. The results of the experiments prove that, although SVM shows decent results in a controlled climate, where its error is 85 percent on CK+, it becomes much less accurate (72 percent) in an uncontrolled environment on FER2013. Comparatively, CNN is doing far better because it achieves 92 and 87 percent accuracy, respectively, in CK+ and FER2013, showing stability to pose, illumination, and noise perturbation. A further analysis showed that CNN achieved better average scores on precision (0.91), recall (0.89), and F1-score (0.90) than SVM, which had precision (0.80), recall (0.77), and F1-score (0.78). Confusion matrices showed that SVM was less able to discriminate between gradients of two classes (fear and sadness), and CNN was more diagonal dominant with less error. Analysis of dataset distributions revealed serious imbalance in FER2013, with happy (8,989 samples) and neutral (6,198 samples) prevailing and affecting the recognition of minority categories, such as disgust (547 samples). The results affirm CNN as a more scalable and consistent FER implementation in real-world applications. Interpretability, computational efficiency, and the problem of dataset imbalance should, in the future, be tackled to provide even more significant performance and applicability in a real environment.