KAN Based Micro-Expression Spotting Network

Tan, Zi Wei (2025) KAN Based Micro-Expression Spotting Network. Final Year Project (Bachelor), Tunku Abdul Rahman University of Management and Technology.

Text Tan Zi Wei - FULL TEXT.pdf Restricted to Registered users only Download (2MB)

Abstract

Micro-expressions, fleeting facial expressions that reveal true emotional states, present significant challenges in the field of affective computing and emotion recognition. This research aims to advance micro-expression spotting by addressing the limitations of existing methodologies that often rely on apex frames and struggle with issues such as dataset limitations and overfitting. A novel framework is proposed, integrating attention mechanisms, intermediate convolutional layers, and the Kolmogorov–Arnold Network (KAN) architecture, which collectively enhance the model's capability to detect subtle expressions in video sequences. By using benchmark datasets, specifically CASME II and CAS(ME)2, the effectiveness of the proposed architecture is evaluated through various performance metrics. The results demonstrate a substantial improvement in detection accuracy and robustness, highlighting the efficacy of leveraging interval frames for micro-expression spotting. This study not only proposes a more effective approach for identifying micro-expressions but also emphasizes the importance of advanced feature extraction techniques and dynamic learning mechanisms in practical applications, including security, mental health monitoring, and interpersonal communication. Ultimately, the findings contribute valuable insights for future research in emotion recognition and pave the way for developing reliable systems capable of capturing complex human emotions.