Triple-Attribute Perceptron Facial Expression Recognition in Real-World Environments

Facial expression recognition (FER) has become a prominent research area due to its wide range of applications, such as human-robot interaction (HRI), driver state monitoring, and medical diagnosis. However, real-world environments pose significant challenges to FER, including occlusion, variations in lighting, and different angles. In this study, a novel triple-attribute perceptron network (TAPNet) is proposed to tackle the limited effectiveness of FER in real-world environments. TAPNet improves FER performance in real-world environments by effectively leveraging triple-attribute facial features from global, local, and critical subregions, thereby fully exploiting the diverse potential information provided by each facial attribute, similar to the human face perception mechanism that extracts both global and regional information. Specifically, the global facial perception (GFP) module emphasizes the most important facial features in the overall face by increasing the number of channels to preserve features and assigning weights to different channels. Additionally, the local facial perception (LFP) and critical facial perception (CFP) modules capture regional feature information from local and critical facial features, respectively, focusing on fine-grained regional features and minimizing interference from irrelevant regions during feature extraction. The experimental results indicate that the proposed TAPNet model achieves an accuracy of 90.77% on the RAF-DB dataset and 65.13% on the AffectNet-7 dataset. Moreover, this model also demonstrates promising FER performance compared to the state-of-the-art approaches on several real-world datasets.