CNLA: Collaborative noisy label adaptive learning for facial expression recognition

Abstract

Existing in-the-wild facial expression recognition (FER) methods rely heavily on predefined labels to achieve high performance. However, in-the-wild FER datasets contain numerous noisy labels, as the uncertainty of facial expressions arises from ambiguous annotations or inter-similarity. Noisy labels provide misleading supervision for learning, leading to decreased generalization. We propose a Collaborative Noisy Label Adaptive Learning (CNLA) method for FER from a new perspective of sample selection to mitigate label inconsistency. CNLA generates perturbed and mixed samples, using Mixed Samples Correction Loss to capture more precise information from various perturbed samples while learning rich representations. Additional information from the perturbed samples is then used for collaborative training, categorizing samples into learnable and relabeled ones. Finally, CNLA constrains the semantic consistency of facial expressions, allowing the model to focus on expression-related regions. Extensive experiments on synthetic noise and original datasets validate the effectiveness of CNLA, demonstrating performance that surpasses state-of-the-art methods.