1.College of Computer Science,Sichuan University,Chengdu 610065,China
2.Guizhou Province Drug Rehabilitation Administration,Guiyang 550003,China
3.Guizhou Women's Compulsory Isolation Drug Rehabilitation Center,Guiyang 550003,China
4.Guizhou Hegu Information Technology Co. ,Ltd. ,Guiyang 550081,China
Show less
文章历史+
Received
Accepted
Published
2026-02-04
2026-03-31
2026-09-28
Issue Date
2026-10-22
PDF (1001K)
摘要
现有基于深度学习的微表情识别方法通常引入面部解剖结构先验以辅助特征建模,但仍存在先验信息参与特征学习不充分、利用方式较为固定,以及缺乏样本自适应调节能力等局限。针对上述问题,本文提出一种基于面部动作编码系统(Facial Action Coding System,FACS)的自适应解剖引导Transformer微表情识别方法(Adaptive FACS-guided Anatomical Transformer for Micro-expression Recognition,AFAT-MER),旨在依据表情强度将FACS先验自适应地融入特征学习与自注意力建模,从而增强模型对微表情关键区域及细微变化的表征能力。首先,利用FACS相关动作单元构建粗粒度面部解剖区域先验,并将其嵌入特征学习过程;其次,在自注意力建模中引入软先验引导,并结合样本表情强度自适应调节先验权重,以提高先验利用的灵活性和样本适应性。实验结果表明,AFAT-MER在多个主流微表情数据集上取得了较好的识别性能,相较基线方法HTNet具有稳定改进,并在与多种代表性方法的比较中表现出一定优势。
Abstract
Although existing deep learning-based micro-expression recognition (MER) methods often incorporate facial anatomical structure priors to facilitate feature modeling, such prior knowledge is typically underutilized, applied in a rigid manner, and lacks sample-level adaptability. To address these limitations, this paper proposes AFAT-MER (Adaptive FACS-guided Anatomical Transformer for Micro-expression Recognition), a novel framework that adaptively integrates Facial Action Coding System (FACS) priors into feature learning and self-attention mechanisms guided by expression intensity. Specifically, the proposed method first constructs coarse facial anatomical region priors inspired by FACS-related Action Units (AUs) and embeds them into the feature learning pipeline. Subsequently, a soft prior guidance mechanism is introduced into the self-attention modeling, where prior weights are dynamically adapted based on expression intensity, thereby enhancing the flexibility and sample-specific suitability of prior utilization. Extensive experiments conducted on several mainstream MER datasets demonstrate that the proposed method achieves competitive recognition performance, yields stable improvements over the HTNet baseline, and exhibits distinct advantages compared with multiple state-of-the-art representative methods.
基于AU的方法试图利用这些先验知识提升识别性能。Zhao等[15]提出的深度区域和多标签学习(Deep Region and Multi-label Learning, DRML)方法通过学习区域特征和AU标签之间的关系来提升检测性能。Li等[14]结合区域适应、多标签学习和最优时序融合进行AU检测,通过多任务学习同时处理多个AUs的检测。近年来,AC2D[16]框架使用自适应约束的自注意力机制结合因果干预来应对AU的细微性和多样性。AUFormer[17]利用Vision Transformer,引入知识专家混合(Mixture-of-Knowledge Expert)机制进行个性化AU学习。这些方法在AU检测任务上取得了不错的效果,其优点在于能够显式利用FACS的解剖-表情映射关系。然而,这些基于AU的方法往往需要额外的AU标注数据,标注成本高昂且耗时,在实际应用中受到限制。此外,这些方法通常采用两阶段设计(先检测AU再分类),可能导致误差传播,且难以实现端到端优化。
YanW J, WuQ, LiangJ, et al. How fast are the leaked facial expressions: The duration of micro-expressions[J]. J Nonverbal Behav, 2013, 37(4): 217-230.
[2]
EkmanP, FriesenW V. Nonverbal leakage and clues to deception[J]. Psychiatry, 1969, 32(1): 88-106.
[3]
PicardR W. Affective computing[M]. Cambridge, Mass: MIT Press, 2000.
[4]
ZhaoG, PietikainenM. Dynamic texture recognition using local binary patterns with an application to facial expressions[J]. IEEE Trans Pattern Anal Mach Intell, 2007, 29(6): 915-928.
[5]
ZachC, PockT, BischofH. A duality based approach for realtime TV-L 1 optical flow[M]//Pattern recognition. Berlin-Heidelberg:Springer,2007: 214-223.
[6]
HeK, ZhangX, RenS, et al. Deep residual learning for image recognition[C]//2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Las Vegas, USA: IEEE, 2016: 770-778.
[7]
HuangG, LiuZ, Van Der MaatenL, et al. Densely connected convolutional networks[C]//2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Honolulu, HI: IEEE, 2017: 2261-2269.
[8]
DosovitskiyA, BeyerL, KolesnikovA, et al. An image is worth 16×16 words: transformers for image recognition at scale[PP/OL]. V1.[2025-12-20].
[9]
WangZ, ZhangK, LuoW, et al. HTNet for micro-expression recognition[J]. Neurocomputing, 2024, 602: 128196.
[10]
EkmanP, FriesenW V. Facial action coding system: A technique for the measurement of facial movement[M]. Palo Alto: Consulting Psychologists Press, 1978.
[11]
TranD, BourdevL, FergusR, et al. Learning spatiotemporal features with 3D convolutional networks [C]//2015 IEEE International Conference on Computer Vision (ICCV). Santiago, Chile:IEEE, 2015: 4489-4497.
[12]
ShaoZ, ChengY, LiF, et al. MOL: Joint estimation of micro-expression, optical flow, and landmark via transformer-graph-style convolution[J]. IEEE Trans Pattern Anal Mach Intell, 2025, 47(10): 8756-8768.
[13]
ZhaoS, LiS, ZhangY, et al. Channel self-attention residual network: Learning micro-expression recognition features from augmented motion flow images [J]. IEEE Trans Affect Comput, 2025, 16(4): 3215-3230.
[14]
LiX, PfisterT, HuangX, et al. A spontaneous micro-expression database: Inducement, collection and baseline [C]//2013 10th IEEE International Conference and Workshops on Automatic Face and Gesture Recognition (FG), 2013: 1-6.
[15]
ZhaoK L, ChuW S, ZhangH G. Deep region and multi-label learning for facial action unit detection [C]//2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016: 3391-3399.
[16]
ShaoZ, ZhuH, ZhouY, et al. Facial action unit detection by adaptively constraining self-attention and causally deconfounding sample[J]. Int J Comput Vis, 2025, 133(4): 1711-1726.
[17]
YuanK, YuZ, LiuX, et al. AUFormer:Vision transformers are parameter-efficient facial action unit detectors[M]//Computer vision-ECCV 2024. Gewerbestrasse,Cham, Switzerland:Springer Nature Switzerland,2024: 427-445.
[18]
LiuS, MaoX, ZhaoS, et al. MER-CLIP: AU-guided vision-language alignment for micro-expression recognition[J]. IEEE Trans Affective Comput, 2025, 16(4): 3028-3042.
[19]
VaswaniA, ShazeerN, ParmarN, et al. Attention is all you need[C]//Advances in Neural Information Processing Systems 30, 2017: 5998-6008.
[20]
CarionN, MassaF, SynnaeveG, et al. End-to-end object detection with transformers[C]//European Conference on Computer Vision. Cham: Springer, 2020: 213-229.
[21]
ZhangK, ZhangZ, LiZ, et al. Joint face detection and alignment using multitask cascaded convolutional networks[J]. IEEE Signal Process Lett, 2016, 23(10): 1499-1503.
[22]
YanW J, LiX, WangS J, et al. CASME Ⅱ: An improved spontaneous micro-expression database and the baseline evaluation [J]. PLoS One, 2014, 9(1): e86041.
[23]
DavisonA K, LansleyC, CostenN, et al. SAMM: A spontaneous micro-facial movement dataset[J]. IEEE Trans Affect Comput, 2018, 9(1): 116-129.
[24]
KingmaD P, BaJ. Adam: A method for stochastic optimization[PP/OL]. V9.(2017-01-30)[2025-12-20].
[25]
ZhangM, YangW, WangL, et al. HFA-Net: Hierarchical feature aggregation network for micro-expression recognition[J]. Complex & Intelligent Systems, 2025, 11(3): 169.