A fluid dynamics-inspired modified reynolds number for microexpression recognition

Abstract Facial expressions exhibit a wide range of emotions. They are regulated by neural mechanisms in the brain that are beyond one’s control. Facial expressions are largely involuntary. The term “micro facial expression” refers to subtle and brief muscle movements that can serve as an indicator of deception. However, existing methods used to detect and analyze facial expressions have certain limitations. These methods focus primarily on spatiotemporal features, while overlooking the speed and intensity of pixel movement. To address these limitations, this study proposes a new method—Reynolds Number-Adapted Microexpression Recognition method—that captures both motion and temporal variations in intensity. A novel equation is formulated by adapting the Reynolds number concept from fluid dynamics to the microexpression recognition task. A pair of onset and apex frames is fed into Recurrent All-Pairs Field Transforms (RAFT), an optical flow technique, to estimate pixel-wise motion. A Reynolds map is then constructed to jointly encode motion magnitude and intensity variation, which is then fed into a Block Division Convolutional Neural Network (BDCNN) to enhance spatial, temporal, and intensity learning. Experiments were conducted using the composite dataset, which is a combination of CASME II and SAMM benchmark datasets. The results showed that the proposed model achieved an accuracy of 86.95% and exhibited improved microexpression recognition.