Selected work

Representative Papers

Cross-Sign Language Transfer Learning Using Domain Adaptation with Multi-scale Temporal Alignment

Aug 17, 2026

This study addresses the challenges of data scarcity and cross-lingual transfer in sign language recognition by proposing a multi-scale temporal relation alignment mechanism based on Temporal Relation Networks (TRN). Integrating TA3N domain adaptation with optical flow features, this approach effectively mitigates domain shift through multi-scale temporal modeling, thereby significantly enhancing RGB modality representations. Experimental results demonstrate that the proposed framework outperforms conventional transfer learning methods in American Sign Language (ASL) recognition tasks. These findings validate the efficacy of RGB-based approaches in low-resource scenarios and provide an efficient solution for cross-lingual sign language recognition, offering a robust pathway to overcome resource limitations in the field.

0 citationsRead paper

Spatio-Temporal Graph Structure Learning for Earthquake Detection

Mar 14, 2025

To address the poor robustness of single-station seismic detection under low signal-to-noise ratio (SNR) conditions—thereby compromising early earthquake warning performance—this paper proposes a spatiotemporal dynamic graph modeling framework for multi-station waveform data. The core innovation is Spectral Structure-Learning Convolution (Spectral SLC), the first method to jointly model static station geographical topology and dynamic seismic wave propagation relationships, overcoming the limitation of conventional Graph Convolutional Networks (GCNs) that rely on fixed graph structures. Building upon this, we design an end-to-end Spatiotemporal Graph Convolutional Network (ST-GCN) that outputs station-level seismic detection probabilities. Evaluated on a real-world earthquake dataset, our method significantly outperforms baseline GCN models: true positive rate (TPR) improves by 12.6%, and false positive rate (FPR) decreases by 41.3%, demonstrating substantial gains in both detection accuracy and robustness through multi-station collaborative inference.

0 citationsRead paper

Recent publications

Latest Papers

Cross-Sign Language Transfer Learning Using Domain Adaptation with Multi-scale Temporal Alignment

Aug 17, 2026

This study addresses the challenges of data scarcity and cross-lingual transfer in sign language recognition by proposing a multi-scale temporal relation alignment mechanism based on Temporal Relation Networks (TRN). Integrating TA3N domain adaptation with optical flow features, this approach effectively mitigates domain shift through multi-scale temporal modeling, thereby significantly enhancing RGB modality representations. Experimental results demonstrate that the proposed framework outperforms conventional transfer learning methods in American Sign Language (ASL) recognition tasks. These findings validate the efficacy of RGB-based approaches in low-resource scenarios and provide an efficient solution for cross-lingual sign language recognition, offering a robust pathway to overcome resource limitations in the field.

0 citationsRead paper

Spatio-Temporal Graph Structure Learning for Earthquake Detection

Mar 14, 2025

To address the poor robustness of single-station seismic detection under low signal-to-noise ratio (SNR) conditions—thereby compromising early earthquake warning performance—this paper proposes a spatiotemporal dynamic graph modeling framework for multi-station waveform data. The core innovation is Spectral Structure-Learning Convolution (Spectral SLC), the first method to jointly model static station geographical topology and dynamic seismic wave propagation relationships, overcoming the limitation of conventional Graph Convolutional Networks (GCNs) that rely on fixed graph structures. Building upon this, we design an end-to-end Spatiotemporal Graph Convolutional Network (ST-GCN) that outputs station-level seismic detection probabilities. Evaluated on a real-world earthquake dataset, our method significantly outperforms baseline GCN models: true positive rate (TPR) improves by 12.6%, and false positive rate (FPR) decreases by 41.3%, demonstrating substantial gains in both detection accuracy and robustness through multi-station collaborative inference.

0 citationsRead paper