Scholarlink

<p>Continuous sign language recognition (CSLR) seeks to transcribe unsegmented sign language videos into gloss sequences without frame-level supervision, presenting persistent challenges in temporal alignment, long-range dependency modeling, and reliable sequence-level generalization. While recent advances have achieved strong performance in high-resource languages, Kurdish sign language (KrdSL) remains largely unexplored due to the absence of sentence-level benchmarks. To address this gap, we introduce KrdSL-1400, the first continuous Kurdish sign language dataset, comprising 1400 annotated video sequences covering 40 linguistically structured sentences performed by seven native signers, providing a standardized benchmark for low-resource CSLR. We propose a hybrid spatio-temporal CSLR framework that combines deep convolutional visual encoding with sequence-aware temporal modeling and a multi-task joint CTC–attention decoding strategy adapted to address alignment uncertainty in CSLR. The CTC objective enforces monotonic alignment, while an entropy-regularized multi-head attention mechanism dynamically emphasizes linguistically salient temporal segments, enabling robust sequence prediction without reliance on pose estimation or handcrafted features. Training dynamics exhibit stable and consistent convergence, with closely aligned training and validation WER curves, indicating reliable model learning. Quantitative evaluation shows that the CTC baseline achieves a WER of 13.5%, which is reduced to 10.5% with single-head attention, while the proposed model attains the best performance, achieving a WER of 9.5%, corresponding to an approximate 30% relative improvement. Employing a multi-dataset evaluation protocol, in which training and testing are conducted independently on the proprietary KrdSL dataset and the large-scale PHOENIX-2014-T benchmark, the proposed model demonstrates dataset-wise robustness, attaining a WER of 13.7% on PHOENIX-2014-T and surpassing recent attention-based and transformer-based CSLR approaches.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Entropy-regularized joint CTC–attention learning for low-resource continuous sign language recognition

  • 期刊论文

  • 发表日期：

    2026年6月20日

查看全文
  • Hanan A. Taher,
  • Subhi R. M. Zeebaree

摘要

Continuous sign language recognition (CSLR) seeks to transcribe unsegmented sign language videos into gloss sequences without frame-level supervision, presenting persistent challenges in temporal alignment, long-range dependency modeling, and reliable sequence-level generalization. While recent advances have achieved strong performance in high-resource languages, Kurdish sign language (KrdSL) remains largely unexplored due to the absence of sentence-level benchmarks. To address this gap, we introduce KrdSL-1400, the first continuous Kurdish sign language dataset, comprising 1400 annotated video sequences covering 40 linguistically structured sentences performed by seven native signers, providing a standardized benchmark for low-resource CSLR. We propose a hybrid spatio-temporal CSLR framework that combines deep convolutional visual encoding with sequence-aware temporal modeling and a multi-task joint CTC–attention decoding strategy adapted to address alignment uncertainty in CSLR. The CTC objective enforces monotonic alignment, while an entropy-regularized multi-head attention mechanism dynamically emphasizes linguistically salient temporal segments, enabling robust sequence prediction without reliance on pose estimation or handcrafted features. Training dynamics exhibit stable and consistent convergence, with closely aligned training and validation WER curves, indicating reliable model learning. Quantitative evaluation shows that the CTC baseline achieves a WER of 13.5%, which is reduced to 10.5% with single-head attention, while the proposed model attains the best performance, achieving a WER of 9.5%, corresponding to an approximate 30% relative improvement. Employing a multi-dataset evaluation protocol, in which training and testing are conducted independently on the proprietary KrdSL dataset and the large-scale PHOENIX-2014-T benchmark, the proposed model demonstrates dataset-wise robustness, attaining a WER of 13.7% on PHOENIX-2014-T and surpassing recent attention-based and transformer-based CSLR approaches.