FACIAL EMOTION RECOGNITION USING DEEP LEARNING AND BI-LSTM

Abstract

Recognizing facial expressions is crucial. component in the development of Human-Computer interaction (HCI) since it makes it possible for the machine to understand and respond to human emotions. The deep learning techniques for FER presented in this article are based on the fusion of a bidirectional long short-term memory network and a convolutional neural network. These methods are meant to be applied in order to create strong emotion classification systems by utilizing temporal and spatial factors. The proposed method takes into account the JAFFE and CK+ datasets. The five basic emotion classes—happy, sad, surprised, angry, and fear—are represented by the images in both sets. The process of preprocessing begins with the detection of faces through the utilization of Multi-task Cascaded Convolutional Networks (MTCNN), followed by the normalizing and scaling of the faces in order to maintain consistency in the input dimensions. Conversely, the CNN component is in charge of retrieving spatial properties from the images, while the BILSTM component is in charge of capturing temporal dynamics that describe face expressions across time. It is clear that such a combination architecture will lead to an increase in recognition accuracy because both types of features are integrated into a single structure. Its accuracy significantly increased when it was trained on the pre-processed datasets for a total of 25 epochs. The model's accuracy during training and validation was 95.2% and 93.1%, respectively, indicating that it can effectively generalize to data that it has never encountered before. The problems related to illumination, position, and the fluctuating intensity of expressions are all efficiently addressed by the combination of CNN and Bi-LSTM.