PDA-AR: Progressive low-light facial expression recognition via multi-level alignment.

Abstract

Facial expression recognition (FER) under low-light conditions faces multiple challenges, such as image noise, blur, and texture loss, and while modern image enhancement methods can improve visual quality, they often optimize for human perception rather than downstream classification, inadvertently introducing artifacts that corrupt subtle discriminative facial features. This article proposes a progressive low-light facial expression recognition framework (PDA-AR), which achieves knowledge transfer from normal-light to low-light images without the need for image enhancement through a multi-level information alignment mechanism and curriculum learning strategy. Specifically, we designed a Learning Sorting Module (LS) that progressively sorts images by brightness, guiding the model to learn from simple to complex. Simultaneously, the model mitigates the domain gap by minimizing the statistical divergence between normal-light and low-light distributions across multiple network stages. This alignment is achieved by optimizing Maximum Mean Discrepancy for feature and semantic embeddings, and utilizing Structural Similarity to enforce consistency in spatial attention maps, ensuring robust knowledge transfer. The experiment was conducted on the RAF-DB and FERPlus datasets, and the results showed that PDA-AR achieved a recognition accuracy of 87.81% under low-light conditions, which is superior to existing mainstream methods and verifies its good generalization ability and practical value (This paper represents an advancement over our previous research (Wang et al., 2024a)).