To address the challenges in continuous sign language recognition, such as the difficulty of effectively integrating RGB features and skeleton features, and the low recognition accuracy of gestures that are visually similar but semantically different, a continuous sign language recognition model combining cross-modal attention and semantic awareness was proposed. Firstly, a dual-stream S3D network was constructed to extract spatiotemporal features from videos and keypoint sequences, respectively, and a cross-modal attention mechanism was used to fuse multimodal information, enhancing feature complementarity and temporal alignment. Next, a language-aware label smoothing strategy was proposed, generating soft labels based on the cosine similarity between tokens calculated from FastText pre-trained word vectors, incorporating semantic priors to distinguish gestures that were visually similar but semantically different. Finally, the entire model was jointly optimized with connectionist temporal classification(CTC) loss, smoothed classification loss, and visual branch auxiliary loss. Multiple comparison experiments were conducted on the Phoenix-2014, Phoenix-2014T, CSL-Daily, and ArabSign datasets. Experimental results show that the word error rates of the proposed method on these four datasets are reduced to 18.2%, 18.8%, 24.3%, and 47.6%, respectively.
WEI Ruo-yan, LIU Xin-qi, ZHU Xiao-qing, LI Xiao-juan.
Continuous Sign Language Recognition with Dual-stream Spatiotemporal Encoding and Semantic Awareness[J].
Science Technology and Engineering, 2026, 26(29): 12698-12713 DOI:10.12404/j.issn.1671-1815.2507441
连续手语识别(continuous sign language recognition,CSLR)是一类将未分割手语视频转换为文本序列(gloss)的关键技术,是连接听障与健听人群的重要桥梁[1-2]。现有连续手语识别方法基于输入数据类型的不同可分为两类:基于RGB特征的方法,融合其他模态的方法。早期研究通常通过卷积神经网络(convolutional neural networks,CNN)提取视频帧的空间特征,并利用循环神经网络(regression neural network,RNN)或长短期记忆网络(long short-term memory,LSTM)进行时序建模。例如,Koller等[3]提出CNN-LSTM混合框架用于捕捉手语动作的时序变化,显著提升了连续手语识别性能。随后,Carreira等[4]提出的I3D模型通过在二维卷积中引入时间维度,实现了对动作时空特征的联合建模,并成为手语识别任务中的重要基准模型。在此基础上,研究者进一步引入Transformer结构以增强模型对长序列依赖关系的建模能力。例如,Camgoz等[5]提出基于Transformer的手语识别与翻译联合框架,在PHOENIX-2014T数据集[6]上取得较好效果。近年来,Ahn等[7]提出SlowFast结构,通过双路径网络分别捕获低频姿态变化和高频手势细节,并结合Transformer实现全局特征融合,进一步提升了手语识别性能。同时,自监督预训练技术也被引入该领域,如Hu等[8]提出的SignBERT模型,通过大规模手语视频预训练增强了模型的特征表达能力和跨场景泛化能力。
尽管基于RGB视觉特征的方法在连续手语识别中取得了显著进展,但RGB视频容易受到光照变化、背景噪声及遮挡等因素的影响,从而导致特征稳定性下降[9-10]。此外,视觉上高度相似但语义不同的手势(visually similar but semantically distinct signs,VISigns)往往难以仅凭RGB特征进行准确区分[10]。为克服上述问题,多模态信息融合逐渐成为近年来CSLR研究的重要方向。这类方法通常通过融合RGB视频、骨骼关键点或光流等信息,实现多模态特征互补。例如,Chen等[11]提出双流网络结构,将RGB视频与关键点热图分别输入S3D网络[12]进行特征提取,并通过特征融合显著提升识别性能,在连续手语识别任务中取得良好效果[13]。此外,SAM-SLR[14]通过构建关节点图并结合图卷积网络(graph convolutional network,GCN)实现骨骼信息与视觉特征的融合,从而增强模型对复杂手语动作的建模能力。Min等[15]提出的VAC模型利用视觉对齐约束融合光流与RGB信息,提高了词元边界检测精度和整体识别效果。
视觉分支承担着对原始RGB视频流进行高效且细致的时空特征提取任务。考虑到连续手语识别中手势存在大量动态变化和空间位移,尤其是视觉上相似但语义不同的手势(VISigns),模型需要具备强大的局部时空建模能力以避免混淆。为此,引入一种改进的可分离三维卷积网络(separable 3D convolution,S3D)作为视觉分支的主干结构,旨在控制计算成本的同时提升模型对手部、面部和身体动态区域的感知能力。与传统的3D卷积不同,S3D通过解耦时间和空间的建模路径,显著减少了参数量和计算开销。例如传统3D卷积在输入通道数为Cin,输出通道数为 Cout,卷积核大小为Kt×Kh×Kw时,其参数量为
Minawaer·AbulaAlifu·Kuerban, XieQina, et al. Review of sign language recognition methods and techniques[J]. Computer Engineering and Applications, 2021, 57(18): 1-12.
TaoTangfei, LiuTianyu. Survey of sign language recognition technology based on expression content and expression features[J]. Journal of Electronics& Information Technology, 2023, 45(10): 3439-3457.
[5]
KollerO, ForsterJ, NeyH. Continuous sign language recognition: towards large vocabulary statistical recognition systems handling multiple signers[J]. Computer Vision and Image Understanding, 2015, 141: 108-125.
[6]
CarreiraJ, ZissermanA. Quo vadis, action recognition? a new model and the kinetics dataset[C]// Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Honolulu, HI: IEEE, 2017: 6299-6308.
[7]
CamgozN C, KollerO, HadfieldS, et al. Sign language transformers: joint end-to-end sign language recognition and translation[C]// Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. Los Angeles, CA: IEEE, 2020: 10023-10033.
[8]
CamgozN C, KadarA, HadfieldS, et al. Neural sign language translation[C]// Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Salt Lake City, UT: IEEE, 2018: 7784-7793.
[9]
AhnJ, JangY, ChungJ S. Slowfast network for continuous sign language recognition[C]// ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing. Seoul: IEEE, 2024: 3920-3924.
[10]
HuH, ZhaoW, ZhouW, et al. Signbert: pre-training of hand-model-aware representation for sign language recognition[C]// Proceedings of the IEEE/CVF International Conference on Computer Vision. Montreal, QC: IEEE, 2021: 11087-11096.
[11]
JiangS, SunB, WangL, et al. Skeleton aware multi-modal sign language recognition[C]// Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. Nashville, TN: IEEE, 2021: 3413-3423.
LiYunan, GengXi, MiaoQiguang. Review of sign language recognition and translation based on computer vision[J]. Computer Engineering&Science, 2025, 42(6): 15-36.
[14]
ZuoR, WeiF, MakB. Natural language-assisted sign language recognition[C]// Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. Vancouver, BC: IEEE, 2023: 14890-14900.
[15]
ChenY, ZuoR, WeiF, et al. Two-stream network for sign language recognition and translation[C]// Proceedings of the 36th Conference on Neural Information Processing Systems (NeurIPS 2022). New Orleans, LA: NeurIPS Foundation, 2022: 17043-17056.
[16]
XieS, SunC, HuangJ, et al. Rethinking spatiotemporal feature learning: speed-accuracy trade-offs in video classification[C]// Proceedings of the European Conference on Computer Vision (ECCV). Munich: Springer, 2018: 305-321.
WangZhuocheng, ZhangJingqiao. Continuous sign language recognition based on 3D hand skeleton data[J]. Journal of Computer-Aided Design & Computer Graphics, 2022, 34(9): 1681-1691.
[19]
MinS, LeeJ, ChoiS, et al. VAC: visual-alignment constrained network for continuous sign language recognition[C]// Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Nashville, TN: IEEE, 2021: 13590-13599.
HuXuezheng, TaoQing, ZhaoMuchao, et al. Gesture recognition method of surface electromyography signals based on transfer learning[J]. Science Technology and Engineering, 2024, 24(12): 5044-5050.
[22]
ZhangX, LiuY, WangH, et al. SML: a skeleton-based multi-feature learning method for sign language recognition[J]. Knowledge-Based Systems, 2024, 301: 112288.
[23]
AlyM, FathiI S. Recognizing American sign language gestures efficiently and accurately using a hybrid transformer model[J]. Scientific Reports, 2025, 15: 20253.
[24]
WeiD, HuH, MaG F. Part-wise graph Fourier learning for skeleton-based continuous sign language recognition[J]. Journal of Imaging, 2025, 11(8): 286.
[25]
LiuY, ChenZ, WangP, et al. A sensing data and deep learning-based sign language recognition approach[J]. Computers & Electrical Engineering, 2024, 118: 109339.
GuoDan, TangShengeng, HongRichang, et al. Survey on sign language recognition, translation and generation[J]. Computer Science, 2022, 49(6): 1-13.
[28]
KollerO, ZargaranO, NeyH, et al. Deep sign: hybrid CNN-HMM for continuous sign language recognition[C]//Proceedings of the British Machine Vision Conference 2016. New York: BMVA Press, 2016: W2587277634.
[29]
HuangJ, ZhouW, ZhangQ, et al. Video-based sign language recognition without temporal segmentation[C]// Proceedings of the AAAI Conference on Artificial Intelligence. New Orleans, LA: AAAI Press, 2018, 32(1): 2062-2067.
[30]
MinY, HaoA, ChaiX, et al. Visual alignment constraint for continuous sign language recognition[C]// Proceedings of the IEEE/CVF International Conference on Computer Vision. Montreal, QC: IEEE, 2021: 11542-11551.
[31]
FayyazsanaviP, AnastasopoulosA, KošeckáJ. Gloss2Text: sign language gloss translation using LLMs and semantically aware label smoothing[J]. arXiv prePrint, arXiv: 2407. 01394, 2024.
LuShaotong, WangChuanxu. Hybrid multi-channel joint learning and dual-branch attention fusion for action recognition[J]. Computer Engineering and Applications, 2025, 61(8): 145-154.
[34]
CuiY, JiaM, LinT Y, et al. Class-balanced loss based on effective number of samples[C]// Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. Long Beach, CA: IEEE, 2019: 9268-9277.
[35]
NeubigG, DervosD, GoyalN, et al. Neural machine translation with extended context[C]// Proceedings of the Conference on Empirical Methods in Natural Language Processing. Brussels: ACL, 2018: 1527-1537.
[36]
KendallA, GalY, CipollaR. Multi-task learning using uncertainty to weigh losses for scene geometry and semantics[C]// Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Salt Lake City, UT: IEEE, 2018: 7482-7491.
[37]
CamgozN C, HadfieldS, KollerO, et al. SubUNets: end-to-end hand shape and continuous sign language recognition[C]// IEEE International Conference on Computer Vision (ICCV). Venice: IEEE, 2017: 3075-3084.
[38]
PuJ, ZhouW, LiH. Iterative alignment network for continuous sign language recognition[C]// Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Long Beach, CA: IEEE, 2019: 4165-4174.
[39]
KollerO, CamgozN C, NeyH, et al. Weakly supervised learning with multi-stream CNN-LSTM-HMMs to discover sequential parallelism in sign language videos[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2019, 42(9): 2306-2320.
[40]
NiuZ, MakB. Stochastic fine-grained labeling of multi-state sign glosses for continuous sign language recognition[C]// European Conference on Computer Vision (ECCV). Glasgow: Springer, 2020: 172-186.
[41]
HeK, ZhangX, RenS, et al. Deep residual learning for image recognition[C]// Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Las Vegas, NV: IEEE, 2016: 770-778.
[42]
PuJ, ZhouW, HuH, et al. Boosting continuous sign language recognition via cross modality augmentation[C]// Proceedings of the 28th ACM International Conference on Multimedia. Seattle, WA: ACM, 2020: 1497-1505.
[43]
HaoA, MinY, HaoA, et al. Self-mutual distillation learning for continuous sign language recognition[C]// Proceedings of the IEEE/CVF International Conference on Computer Vision. Montreal, QC: IEEE, 2021: 11303-11312.
[44]
ChengK L, YangZ, ChenQ, et al. Fully convolutional networks for continuous sign language recognition[C]// European Conference on Computer Vision. Cham: Springer, 2020: 697-714.
[45]
ZhuQ, LiJ, YuanF, et al. Continuous sign language recognition based on motor attention mechanism and frame-level self-distillation[J]. Machine Vision and Applications, 2025, 36(1): 1-7.
[46]
LuqmanH. ArabSign: a multi-modality dataset and benchmark for continuous Arabic sign language recognition[J]. arXiv prePrint, arXiv: 2210. 03951, 2022.