华中科技大学学报（自然科学版）2025，Vol.53Issue(5)：65-70,6.DOI:10.13245/j.hust.250043
基于手语单词约束的连续手语识别网络
Continuous sign language recognition network based on sign language word constraints
摘要
Abstract
Aiming at the problem that existing continuous sign language recognition methods rely on connectionist temporal classification(CTC)loss for sequence alignment,but they struggle to address intra-class variations and inter-class similarities caused by demonstrator habits and contextual switches,a sign language word-constrained recognition network was proposed,including word existence constraints and word count constraints.The former explicitly emphasized and suppressed target words through average pooling of the top K high-probability time points and encoding of existence labels.The latter combined temporal attention and global average pooling to enhance the ability to capture temporal information.Experiments on the PHOENIX-2014,PHOENIX-2014T,and CSL datasets show that the proposed method achieves word error rates of 21.1%,21.6%and 1.5%on the test sets,respectively,verifying its effectiveness and generalization ability.关键词
计算机视觉/连续手语识别/单词约束/多标签分类/辅助约束Key words
computer vision/continuous sign language recognition/word constraints/multi-label classification/auxiliary constraint分类
信息技术与安全科学引用本文复制引用
侯永宏,闫祉君,尹文杰,周文..基于手语单词约束的连续手语识别网络[J].华中科技大学学报（自然科学版）,2025,53(5):65-70,6.基金项目
国家重点研发计划资助项目(2023YFC2411103). （2023YFC2411103）
