多環境下烏爾都字母唇語識別系統

// Lip reading systems for Urdu alphabets in diverse environments

本研究針對烏爾都語唇讀資料集不足問題，推出ULRA資料集並採用先進資料擴增技術｜評估三種深度神經網路模型，LipNet基礎的2D-CNN在未知資料上達81.97%最高準確率，且精確度、召回率與F1分數均優異｜混合型2D_3D-CNN展現更佳泛化能力，結果驗證ULRA資料集有效支持烏爾都語唇讀技術進步。

// This study addresses the lack of Urdu lip reading datasets by introducing the ULRA dataset and applying advanced data augmentation. It evaluates three deep neural network models, including a LipNet-based 2D-CNN, a Hybrid 2D_3D-CNN, and a baseline 3D-CNN, across familiar and unfamiliar data environments. The LipNet-based 2D-CNN achieved the highest accuracy (81.97%) and superior precision, recall, and F1-score on unknown data, while the Hybrid model showed better generalization. These results highlight the effectiveness of the ULRA dataset and support further advancements in Urdu lip reading technology.

Published 2026年6月14日