摘要
为协助听障人士改善日常沟通状况,提出了一种基于改进SlowFast架构的连续手语识别模型.该模型通过慢速路径和快速路径分别提取手语视频的空间特征和动作特征,并引入双向特征融合加强两个路径的信息交互.此外,利用动作参数A和移动参数M捕获手势变化幅度和频率,优化模型性能.实验结果表明,所提方法在PHOENIX2014和CSL-Daily测试集上的WER分别为18.8%和28.2%,相比其他方法在手语识别的准确率上有明显改进.
In order to improve daily communication for the deaf,a continuous sign language recognition model based on improved SlowFast architecture is proposed.The model extracts spatial and temporal features from sign language videos through the slow and fast pathways.The bidirectional feature fusion is introduced to enhance information interaction between the two paths.Additionally,motion parameter A and mobility parameter M are utilized to capture gesture amplitude and frequency,further optimizing model performance.Experimental results demonstrate that the proposed method achieves WER of 18.8%and 28.2%on the PHOENIX14 and CSL-Daily test datasets,significantly outperforming existing approaches in continuous sign language recognition.
作者
蒋敏敏
JIANG Minmin(Putian University,Putian 351100,China)
出处
《通化师范学院学报》
2025年第10期34-40,共7页
Journal of Tonghua Normal University
基金
福建省中青年教师教育科研项目(科技类)(JAT220299)。
关键词
连续手语识别
SlowFast
动作参数
移动参数
continuous sign language recognition
SlowFast
motion parameter
mobility parameter
