Frequency-Guided Expert Modulation for Noisy-Label Facial Expression Recognition

Abstract

Facial expression recognition in the wild is challenged by both noisy supervision and degraded visual evidence: subtle expression cues must be interpreted under blur, contrast changes, image noise, and annotator disagreement. Existing noisy-label FER methods mainly regulate samples, labels, or attention, while frequency information is rarely used to adapt the semantic representation itself. We propose Frequency-Guided Expert Modulation (FARM-FER), which treats local and global frequency descriptors as a control signal rather than an additional classifier input. A joint Haar-DWT and radial-FFT context guides soft routing among nonlinear experts and channel-wise affine recalibration of the semantic feature, while a learned gate combines the two corrections before a lightweight classifier predicts the expression from the refined representation. Across RAF-DB, FER+, and AffectNet under symmetric label noise, with additional evaluations under class-dependent label noise on RAF-DB and native crowd-label ambiguity on FER+, FARM-FER consistently improves matched baselines. At 30% symmetric noise, FARM-FER reaches 89.18% accuracy on RAF-DB, with a 1.6% performance gain over the matched Swin-Tiny baseline; the gains also hold in a controlled ResNet18 reimplementation and in class-sensitive AffectNet evaluation. Measured cost analyses show only modest parameter and FLOP overhead, supporting a lightweight yet effective design in terms of model size and arithmetic cost for noisy-label FER.