PCD-Net: prior-context collaborative learning for robust facial expression recognition under noisy labels

Abstract Noisy annotations and class imbalance commonly coexist in real-world facial expression recognition (FER), making hard but correctly labeled minority-class samples difficult to distinguish from unreliable training samples. Existing noisy FER approaches mainly improve robustness through sample selection or re-weighting, while the representation diversity used for collaborative denoising and the distinction between sample reliability and class difficulty remain insufficiently explored. To address these issues, we propose a prior-context collaborative learning network with dynamic compensation (PCD-Net). PCD-Net employs two heterogeneous encoders with complementary representation biases. The prior branch uses an Attention-based Hierarchical Part Representation module to learn multiple spatially selective facial responses, whereas the context branch employs Semantic-Guided Feature Alignment to calibrate global semantic and salient contextual cues. For noise-robust optimization, we introduce a Dynamic Relative Beta Mixture Model, which estimates sample-level reliability from prediction-label alignment and separately models class-relative learning difficulty through temporally bounded dynamic compensation. Extensive experiments on RAF-DB, FERPlus, and AffectNet under symmetric label noise, together with asymmetric and controlled instance-dependent noise evaluations on RAF-DB, demonstrate that PCD-Net is particularly effective under moderate-to-severe corruption. Additional class-balanced, ablation, branch-complementarity, inference-strategy, and computational analyses further characterize its robustness and efficiency trade-offs.