Enhancing Occlusion-robustness for Micro-expression Recognition
Year: 2026
Type: article
Abstract: Micro-expressions reveal the true feelings of humans and can benefit a wide range of applications as they provide important cues pertaining to affective computing and psychology. With the recent advancement of deep learning, micro-expression recognition approaches have made great progress and achieved impressive performances. Although automatic micro-expression recognition has practical importance in various fields, existing micro-expression recognition approaches cannot generalize well to occluded face images, which are common cases in real-world scenarios. In this paper, we present a novel approach aiming to recognize the micro-expression under severe facial occlusions. Specifically, we propose a Micro-expression Localization Network (MELN) to discover the decisive patches within a micro-expression image for recognition, ignoring the irrelevant region to reduce the impact of noise induced by facial occlusions. Additionally, to overcome the data-insufficient issue of limited micro-expression datasets, we introduce Cross-subject Adaptive Mixing (CAMixing), a novel mixing-based data augmentation approach to synthesize advantageous micro-expression samples that help to improve the robustness of the recognition model. To the best of our knowledge, this is the first work to discuss the performance degradation of micro-expression recognition methods with the presence of severe facial occlusions and propose a micro-expression recognition approach aiming to improve the recognition robustness. Extensive experimental results show the effectiveness and robustness of our approach, manifesting that our proposed method outperforms the state-of-the-art approaches on occluded micro-expressions.
Source: IEEE Transactions on Multimedia
Language: English
Topic: Emotion and Mood Recognition
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
Estimated APC paid: $2,645
OpenAlex ID: W7203948552