Robust Facial Emotion Recognition System via De-Pooling Feature Enhancement and Weighted Exponential Moving Average

Facial expression recognition (FER) in an uncontrolled environment presents a formidable challenge in affective computing and human-machine interaction domains. Existing FER models fail to generalize due to the innate expression nature of the intra-class separability and inter-class compactness. Although most state-of-the-art FER research focuses on more advanced frameworks and well-separated discriminative loss functions, further exploration is needed to obtain high-quality expression features. This paper proposes a robust feature enhancement approach for FER by integrating the De-pooling Feature Enhancement and Weighted Exponential Moving Average (WEMA) of Stochastic Gradient Descent (SGD). The proposed method utilizes De-Pooling Feature Enhancement to capture and enhance the high-quality expression features, while WEMA of SGD optimizes the training process for improved stability and convergence. Our extensive experimental analysis of benchmark datasets signifies that our proposed method prevails over state-of-the-art methods, achieving superior accuracy performances of 86.43% on FER2013, 94.97% on FERPlus, 94.71% on RAF-DB, as well as 80.62% and 72.90% on AffectNet, of 7 class and 8 class, respectively.