EFFICIENT METHODS FOR FACIAL MICRO-EXPRESSIONS DETECTION & CLASSIFICATION

Abstract

Micro-expressions (ME) are a form of facial expression that lasts just a few milliseconds. Due to its use in revealing implicit human intentions, particularly in high-stakes situations, ME detection and classification is a burgeoning field of research. In recent studies, it has been found that the apex frame contains maximum facial movement conveying expression information. But they last for very few milliseconds, hence detecting the apex frame and classifying the ME is a very challenging task. This paper addresses this problem by proposing a training-free method for ME detection and an ME classifier using hybrid features. For ME detection, video frames are preprocessed to remove head movement and align faces. Then, from five local facial regions, Local Binary Patterns (LBP) are extracted. These regions are selected to reduce the effects of head eye blinks and changing eye gaze. Then the apex frame is detected using the Frobenius norm on extracted LBP features. For ME classification, hybrid features: Frame Difference (FD), Optical Flow Magnitude (OF-Mag), and Optical Flow Histogram of Oriented Gradients are extracted from regions around facial landmarks and are used to classify micro-expressions. These regions are selected to reduce the effects of head movement, eye blinks, and changing eye gaze. The proposed methods are tested on publicly available benchmark datasets CASME II. On the CASME II, the ME detection method achieved 0.1137 Normalized Mean Absolute Error (NMAE), 0.0720 Normalized Standard Error (NSE), and the ME classifier achieved 67.07 percent and 0.69, accuracy and a f1-score respectively. The results reveal that the proposed methods excelled the existing methods.