A deep learning framework for micro-expression recognition via multi-feature fusion

In order to enhance the quality of inclusive education and promptly assess the knowledge comprehension of special education students through their facial expression changes during lectures, it is essential to recognize micro-expressions. However, the subtle variations in micro-expressions pose challenges to detection. To enhance micro-expression recognition, a learning method combining Convolutional Neural Networks (CNN) and Long Short-Term Memory (LSTM) networks for spatiotemporal feature extraction and recognition is proposed. Initially, CNN extracts spatial features from the dataset using a dual-channel configuration, followed by LSTM to process temporal features. The sequence is looped with a fixed frame count. Experiments are performed on a test dataset to compare the recognition efficiency of the proposed algorithm with several other algorithms using the existing database. The results demonstrate that the proposed method achieves a comprehensive micro-expression recognition efficiency of 76.24%, representing a significant improvement.