Abstract

Facial expression recognition (FER) is crucial for human-computer interaction, emotion analysis, and psychological studies. Traditional FER models face challenges in handling noisy video data caused by lighting variations, occlusions, and facial distortions, which degrade recognition accuracy. A noise-aware adaptive weighting model is introduced to address these limitations. The proposed approach leverages a spatiotemporal convolutional neural network (CNN) combined with an attention-based adaptive weighting mechanism that dynamically adjusts the contribution of frames based on their noise level. The model processes video sequences by first extracting frame- level features using a CNN. A noise detection module estimates the noise level in each frame, and an adaptive weighting mechanism assigns higher weights to less noisy frames. The weighted features are then passed through a recurrent neural network (RNN) to capture temporal dependencies. Experimental results on the CK+, Oulu- CASIA, and AFEW datasets demonstrate that the proposed model achieves 92.8% accuracy, outperforming existing FER methods such as Temporal CNN and LSTM-based models. The adaptive noise-aware weighting mechanism enhances robustness by reducing the impact of noisy frames, leading to improved expression recognition.

Authors

V. Porkodi
Sivas University of Science and Technology, Turkey

Keywords

Facial Expression Recognition, Adaptive Weighting, Noise-Aware, Spatiotemporal CNN, Recurrent Neural Network