On Applying Label Distribution Learning for Noisy Label Face Emotion Recognition

초록

Face emotion detection task is challenging due to the intricate nature of human facial expressions, which involve ambiguous features and subtle variations in different regions, presenting significant difficulties. The complexity gets further complicated by the recognition of many emotions inside a single facial expression, resulting in the recognition of facial emotions as a classification problem with noisy labels. In this study, we introduce a network that addresses the problem of noisy influences in facial emotion recognition (FER) by utilizing an attention map consistency-based approach and a label distribution generator. The experimental findings demonstrate that our suggested approach exhibits superior performance compared to earlier state-ofthe-art models that were trained on noisy-labeled data.