An Analysis of the Effectiveness of KAN and CNN Algorithms for Human Facial Emotion Classification

Facial expression-based emotion recognition has various applications, including security, healthcare, human-computer interaction, and behavioral analysis. This study analyzes the effectiveness of the Kolmogorov Arnold Networks and Convolutional Neural Networks algorithms in classifying human facial expressions of emotion using the OSEMN. Experimental data were obtained from the FER-2013 dataset, which consists of 35,887 facial expression images categorized into seven primary emotions: happy, sad, angry, fearful, disgusted, surprised, and neutral. The CNN model was designed with four convolutional layers, while the KAN model used three convolutional layers and a B-spline-based approach to handle non-linear transformations. Evaluation was based on accuracy, precision, recall, F1-score, and computational efficiency. The results showed that CNN achieved higher accuracy but tended to overfit, particularly on emotion classes with imbalanced data distribution. On the other hand, KAN demonstrated more stable performance with lower computational resource consumption, making it more efficient for systems with limited power and memory. CNN was selected for its superior pattern recognition capability, while KAN was chosen due to its efficiency in resource-constrained environments. From the comparison, CNN performed better in detecting complex expressions, while KAN was more optimal in processing efficiency and classification stability. The choice of the most suitable algorithm depends on the specific needs of the system—whether prioritizing high accuracy (CNN) or computational efficiency (KAN). This study is expected to provide insights into the development of more adaptive and efficient deep learning-based emotion recognition systems for practical applications such as mobile devices, healthcare monitoring, and smart surveillance.