MSAFNet: a novel approach to facial expression recognition in embodied AI systems

In embodied artificial intelligence (EAI), accurately recognizing human facial expressions is crucial for intuitive and effective human-robot interactions. We introduce multi-scale attention and convolution-transformer fusion network, a deep learning framework tailored for EAI, designed to dynamically detect and process facial expressions, facilitating adaptive interactions based on the user's emotional state. The proposed network comprises three distinct components: a local feature extraction module that utilizes attention mechanisms to focus on key facial regions, a global feature extraction module that employs Transformer-based architectures to capture comprehensive global information, and a global-local feature fusion module that integrates these insights to enhance facial expression recognition accuracy. Our experimental results on prominent datasets such as FER2013 and RAF-DB indicate that our data-driven approach consistently outperforms existing state-of-the-art methods.