Spontaneous Micro-facial Expression Detection using Attention-based Convolutional Gated Recurrent Neural Networks with RMSProp Optimization

Abstract

Introduction: Facial micro expressions have a significant role in communicating hidden emotions, offering profound implications in multiple areas such as psychology, cybersecurity, and the study of human-computer interaction. However, recognizing micro-expressions is tough because of their transient nature. Furthermore,Micro expressions are significantly shaped by the visual characteristics of the face and the interactions among its various sub-regions. Recent advances in computer vision have improved self-supervised learning.

 Traditional CNNs for vision disorders learn only from whole photos or videos and cannot consistently distinguish facial micro expressions. Some existing CNN models might  lack advanced attention mechanisms to effectively prioritize subtle micro-expression features within the complex facial data. They may leading to potential misinterpretation of expressions.

Objectives: This research offers the Attention Convolutional Gated Recurrent Neural Network with RMSProp (ACGRNN- RMSProp) classifier aims to achieve high accuracy and robustness in micro facial expression recognition. The attention mechanism sharpens the model's emphasis on important facial regions, while convolutional layers gather spatial features, and GRUs effectively capture temporal dynamics. The RMSProp optimizer ensures efficient and stable training, making this approach well-suited for the complex task of recognizing subtle and fleeting micro expressions.. This suggested approach is applied to the FER2013 dataset and compared to conventional micro expression recognition baseline methods.

Methods: The FER2013 dataset is initially loaded and processed. Facial landmarks in this dataset are detected using a Multi-Task Cascaded Convolutional Neural Network, after which features are extracted from the processed data using EfficientNetB3 and InceptionResNetV2. The features obtained from extraction are then fused and subsequently classified using an Attention Convolutional Gated Recurrent Neural Network with RMSProp optimization.

Results:This hybrid model demonstrates exceptional classification performance, achieving precision, recall, and F1-score of 94.5% each, along with an accuracy of 96.43%. Furthermore, it includes additional metrics such as the Matthews Correlation Coefficient (MCC), Receiver Operating Characteristic(ROC) Curve in this field.

Conclusions:This study showcases enhanced recognition accuracy when compared to a range of leading models in the field,eliminated the requirement for system preprocessing procedures such as contrast enhancement, gradient operators, or dimensionality reduction for accurate results.Leverage the inherent relationships between face identification and other face analysis tasks to enhance future performance.