Micro-expression Recognition Based on Apex Frame Using Deep Learning

Micro-expression Recognition Based on Apex Frame Using Deep Learning

연구 분야: Artificial Intelligence

학회: 2023 19th International Conference on Natural Computation, Fuzzy Systems and Knowledge Discovery (ICNC-FSKD)

Emotions reflect human’s inner feelings. The precise recognition of emotions is the key to social communication and human-computer interaction. Micro-expression (ME) cannot be completely suppressed, and it can reveal human’s real emotional state. Therefore, it is an important cue for emotion recognition. However, due to the rapid and subtle nature of ME, extracting efficient features from ME sequences is tricky. ME recognition is still a challenging problem. Recently, deep learning based methods have been proposed in the field of pattern recognition and have achieved promising recognition performance. In this paper, we propose a ME recognition method based on apex frame using deep learning models to extract deep features and perform recognition. Firstly, we introduce deep learning based face detection model, named RetinaFace, to perform face detection and pre-processing operations on apex frames. Secondly, due to the limited availability and imbalance of ME data, we propose to use Deep Convolutional Neural Networks (DCNNs) and transfer learning for ME features extraction and recognition. By introducing a DCNNs based face recognition model called FaceNet, this study transfer the prior and learned knowledge in the field of face recognition to the field of ME recognition, thus reducing the training process and improving the performance of ME recognition. Experiments conducted on CASME II prove that our proposed framework is feasible and achieves good recognition performance. Furthermore, we investigate the impact of the number of apex frames on ME recognition, providing clues to address the problem of limited samples in deep learning-based ME recognition. Our work complements deep learning-based methods for emotion recognition and ME recognition, and demonstrates the powerful capabilities and promising potentialities of deep learning frameworks in ME recognition.