QUANTITATIVE ANALYSIS OF MICRO-EXPRESSION PHENOTYPES VIA CUMULATED OPTICAL FLOW AND LBP-TOP UNDER VARYING ENVIRONMENTAL CONSTRAINTS

Abstract

Micro-expression recognition is a challenging problem in affective computing because the facial muscle movement is subtle and transient. This paper proposes a hybrid framework that combines a Cumulated Optical Flow Vector (COFV) to capture very subtle inter-frame motion and Local Binary Patterns (LBP-TOP) to encode spatiotemporal texture representation. COFV is used to for motion estimation under illumination changes, while LBP-TOP preserves appearance dynamics.The COFV method captures the cumulative motion patterns over consecutive frames and enhances discriminability in micro-expression analysis. LBP-TOP encodes dynamic texture variations, complementing motion-based features. The extracted feature sets are fused to result in a hybrid representation and, consequently, superior recognition performance. Experimental results show that our approach outperforms traditional optical flow and LBP-based methods through improved robustness against illumination variations and noise. The method was evaluated on the CASME II dataset and achieved 85.2% accuracy, 84.1% precision, 83.3% recall, and 83.7% F1-score, outperforming only optical flow based as well as the LBP-TOP based baseline models.The proposed method provides a structured and reliable feature extraction pipeline for micro-expression recognition and is very suitable for psychology, security, and human-computer interaction applications.