1 Key Laboratory of Advanced Manufacturing Technology of Ministry of Education, Guizhou University, Guiyang 550025, China
2 Guizhou Province International Science & Technology Cooperation Base of Electronic Information and Intelligent Applications, Guizhou University, Guiyang 550025, China
Show less
文章历史+
Received
Published
Revised
2025-04-18
2026-04-18
2025-10-17
Online Date
2026-06-08
PDF (4975K)
摘要
人脸属性代表人脸图像中人类可理解的视觉特性,对人脸属性的准确识别是具身机器人与人交互的基础。为了能够同时准确感知人脸表情、五官特征、肤色、年龄等人脸属性,提出了基于状态空间模型的深度多任务学习人脸属性识别方法(state space model-based deep multi-task learning for face attribute recognition,SSM-FAR)。首先,该模型利用ConvNext编码图像局部特征的能力,促使骨干网络充分编码图像细粒度特征以识别面部属性特征;利用特征融合模块以增强模型对多尺度和复杂度特征的识别能力;之后,设计了基于状态空间模型的特征增强方法,弥补骨干网络编码全局特征能力的不足。在多任务人脸属性识别任务中,该方法实现了对人脸的表情以及其余40种面部属性的准确分类。实验结果表明,SSM-FAR在CelebA、LFW-A、RAF-DB、AffectNet数据集的平均准确率分别达到了91.70%、86.73%、91.23%、66.51%,优于所比较的代表性方法。
Abstract
Face attributes are represented as human-interpretable visual characteristics in facial images, and the accurate recognition of facial attributes is regarded as the foundation for human-robot interaction in embodied robots. To enable the simultaneous recognition of facial expressions, facial features, skin tone, age, and other facial attributes, a state space model-based deep multi-task learning method for face attribute recognition (SSM-FAR) was proposed. Firstly, ConvNext's capability to encode local image features was leveraged in SSM-FAR, so that fine-grained image features were fully encoded by the backbone network for facial attribute recognition. In addition, a feature fusion module was employed to enhance the model's ability to recognize multi-scale and complex features. Secondly, a feature enhancement module based on the state space model was utilized to compensate for the backbone network's inadequacy in encoding global features. Our method reached accurate classification of facial expressions and the remaining 40 facial attributes in multi-task for face attribute recognition. Experimental results show that average accuracy rates of 91.70%, 86.73%, 91.23% and 66.51%are achieved on the CelebA, LFW-A, RAF-DB, and AffectNet datasets, and that SSM-FAR outperforms all compared state-of-the-art methods.
TANG Yao, XU Biao, YANG Guan-ci.
State Space Model-based Deep Multi-task Learning for Face Attribute Recognition Method[J].
Science Technology and Engineering, 2026, 26(11): 4750-4757 DOI:10.12404/j.issn.1671-1815.2502781
以状态空间模型(state space model, SSM)作为骨干的Mamba语言模型[20]因其具备对长序列进行高效且准确的推理能力而备受关注。围绕同时准确将人脸表情、五官特征、肤色、年龄等人脸属性分类,从利用人脸图像中相距较远的两个像素之间的联系建立长距离依赖关系的角度,将ConvNeXt与Mamba融合,提出基于状态空间模型的深度多任务学习人脸属性识别方法,为人脸表情及多种面部属性的准确分类提供新方法。
1 基于状态空间模型的深度多任务学习人脸属性识别方法
1.1 算法设计
考虑到ConvNeXt[22]在局部特征提取方面具有良好的性能,而Mamba网络在全局上下文理解方面具有优势。为了能够同时准确将人脸表情、五官特征、肤色、年龄等人脸属性分类,将ConvNeXt与Mamba融合,提出基于状态空间模型的深度多任务学习人脸属性识别方法(state space model-based deep multi-task learning for face attribute recognition, SSM-FAR)。SSM-FAR的网络架构如图1所示,主干网络由局部特征提取模块、特征融合模块、特征增强模块构成。SSM-FAR算法流程如下。
XuB, YangG C. Interpretability research of deep learning: a literature survey[J]. Information Fusion, 2025, 115: 102721.
[2]
SangH, JiangR, WangZ, et al. Scene augmentation methods for interactive embodied AI tasks[J]. IEEE Transactions on Instrumentation and Measurement, 2023, 72: 1-11.
XuJie, ZhongYong, WangYang, et al. Facial attribute estimation and expression recognition based on contextual channel attention mechanism[J]. Journal of Computer Applications, 2025, 45(1): 253-260.
[5]
ChenS, ZhuX, YanY, et al. Identity-aware contrastive knowledge distillation for facial attribute recognition[J]. IEEE Transactions on Circuits and Systems for Video Technology, 2023, 33(10): 5692-5706.
LiuYingfang, WangSong, MaYatong. Face attribute recognition combining feature fusion and task grouping[J]. Computer Engineering, 2023, 49(11): 211-219.
[8]
SutarM B, AmbhaikarA. Formicary swarm optimized deep CNN for facial emotion recognition from human facial expressions[J]. Biomedical Engineering: Applications, Basis and Communications, 2024, 36(5): 2450029.
[9]
HiroyaK, KoichiI, TakafumiA. Face attribute estimation using multi-task convolutional neural network[J]. Journal of Imaging, 2022, 8(4): 105.
[10]
LiuZ, LuoP, WangX, et al. Deep learning face attributes in the wild[C]// Proceedings of the IEEE International Conference on Computer Vision. Santiago: Institute of Electrical and Electronics Engineers Inc., 2015: 3730-3738.
[11]
Learned-MillerE, HuangG B. Labeled faces in the wild: a survey[M]// Advances in Face Detection and Facial Image Analysis. Cham: Springer International Publishing, 2016: 189-248.
[12]
HanH, JainA K, WangF, et al. Heterogeneous face attribute estimation: a deep multi-task learning approach[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2018, 40(11): 2597-2609.
[13]
MaoL, YanY, XueJ, et al. Deep multi-task multi-label CNN for effective facial attribute classification[J]. IEEE Transactions on Affective Computing, 2022, 13(2): 818-828.
[14]
YangY, HuL, ZuC, et al. CL-TransFER: Collaborative learning based transformer for facial expression recognition with masked reconstruction[J]. Pattern Recognition, 2024, 156: 110741.
[15]
DosovitskiyA, BeyerL, KolesnikovA, et al. An image is worth 16×16 words: transformers for image recognition at scale[J]. arXiv Preprint, arXiv: 2010. 11929, 2021.
[16]
MaF, SunB, LiS. Facial expression recognition with visual transformers and attentional selective fusion[J]. IEEE Transactions on Affective Computing, 2023, 14(2): 1236-1248.
[17]
XueF, WangQ, GuoG. TransFER: Learning relation-aware facial expression representations with transformers[C]// Proceedings of the IEEE International Conference on Computer Vision. Montreal: Institute of Electrical and Electronics Engineers, 2021: 3581-3590.
[18]
QinL, WangM, DengC, et al. SwinFace: a multi-task transformer for face recognition, expression recognition, age estimation and attribute estimation[J]. IEEE Transactions on Circuits and Systems for Video Technology, 2024, 34(4): 2223-2234.
[19]
QinL, WangM, LiuX, et al. Faceptor: a generalist model for face perception[C]// European Conference on Computer Vision. Cham: Springer Nature Switzerland, 2024: 240-260.
[20]
LiS, DengW, DuJ. Reliable crowdsourcing and deep locality-preserving learning for expression recognition in the wild[C]// 30th IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2017. Honolulu:Institute of Electrical and Electronics Engineers, 2017: 2584-2593.
SunShouyu, WuFengjiao, LuoZijiang, et al. Multi-tasking face attribute estimation algorithm based on embedded system[J]. Science Technology and Engineering, 2022, 22(8): 3228-3235.
[23]
GuA, DaoT. Mamba: linear-time sequence modeling with selective state spaces[J]. arXiv Preprint,arXiv: 2312.00752, 2024.
[24]
LinT, GoyalP, GirshickR, et al. Focal loss for dense object detection[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2018, 42(2): 318-327.
[25]
LiuZ, MaoH, WuC, et al. A ConvNet for the 2020S[C]// Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition. New Orleans: IEEE Computer Society, 2022: 11966-11976.
[26]
LiL, SunQ, ZhaoL, et al. Face Mamba:a facial emotion analysis network based on VMamba[C]// 2024 7th International Conference on Machine Learning and Natural Language Processing (MLNLP). Chengdu: IEEE, 2024: 1-5.
[27]
MollahosseiniA, HasaniB, MahoorM H. AffectNet: a database for facial expression, valence, and arousal computing in the wild[J]. IEEE Transactions on Affective Computing, 2019, 10(1): 18-31.
YangGuanci, XuBiao, LuoKexin, et al. A review of knowledge graph techniques: construction, reasoning and typical application[J]. Journal of Guizhou University (Natural Sciences), 2025, 42(2): 1-10
LiZhongyi, YangGuanci, LiYang, et al. Visual privacy behavior recognition and protection system based on image semantics for social robot[J]. Journal of Computer-Aided Design & Computer Graphics, 2020, 32(10): 1679-1687.