Scholarlink

<p>The lack of a comprehensive benchmark dataset has limited research on right-to-left scripts, especially Persian text detection and recognition. Using state-of-the-art (SOTA) techniques on current Latin or multilingual benchmark datasets may lead to poor detection or recognition performance in Persian scripts. This paper addresses the research gap in Persian text detection and recognition by introducing a comprehensive Persian (Farsi) text detection and recognition (PTDR) dataset. PTDR contains challenging images in the wild and documents, including scene images captured from indoor and outdoor environments. The documents contain historical, social, and literary text instances. We also propose PTDR-Synth, the largest synthetic Persian text dataset, consisting of more than 200,000 synthetic cropped word images. It can be used as a pre-training dataset and a more reliable resource for scene text recognition approaches. To this end, we use standard evaluation metrics to conduct extensive quantitative and qualitative assessments of SOTA text detection (three models) and recognition (seven models) techniques on the proposed dataset. Our experimental results demonstrate that leveraging only pre-trained SOTA text detection models on multilingual text datasets like ICDAR17-MLT can still perform well on the detection set of the PTDR dataset without further training or fine-tuning. However, we show that the SOTA recognition models’ performance decline on the recognition test set of the prepared PTDR. In addition, we show that the SOTA scene text recognition models perform better on Persian scripts when trained on the large synthetic and real-world Persian text datasets that have been prepared. We also show that model with only trained on scene images are generalizable on unseen Persian document images without further fine-tuning or training. The proposed dataset, trained models, and the code will be publicly available at <a href="https://github.com/zobeirraisi/PTDR">https://github.com/zobeirraisi/PTDR</a>.</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

PTDR: A Real-World and Synthetic Benchmark Dataset for Persian Scene and Document Text Detection and Recognition

  • 期刊论文

  • 发表日期：

    2025年7月15日

查看全文
  • Zobeir Raisi,
  • Valimohammad Nazarzehi Had,
  • Esmaeil Sarani,
  • Raosul Damani

摘要

The lack of a comprehensive benchmark dataset has limited research on right-to-left scripts, especially Persian text detection and recognition. Using state-of-the-art (SOTA) techniques on current Latin or multilingual benchmark datasets may lead to poor detection or recognition performance in Persian scripts. This paper addresses the research gap in Persian text detection and recognition by introducing a comprehensive Persian (Farsi) text detection and recognition (PTDR) dataset. PTDR contains challenging images in the wild and documents, including scene images captured from indoor and outdoor environments. The documents contain historical, social, and literary text instances. We also propose PTDR-Synth, the largest synthetic Persian text dataset, consisting of more than 200,000 synthetic cropped word images. It can be used as a pre-training dataset and a more reliable resource for scene text recognition approaches. To this end, we use standard evaluation metrics to conduct extensive quantitative and qualitative assessments of SOTA text detection (three models) and recognition (seven models) techniques on the proposed dataset. Our experimental results demonstrate that leveraging only pre-trained SOTA text detection models on multilingual text datasets like ICDAR17-MLT can still perform well on the detection set of the PTDR dataset without further training or fine-tuning. However, we show that the SOTA recognition models’ performance decline on the recognition test set of the prepared PTDR. In addition, we show that the SOTA scene text recognition models perform better on Persian scripts when trained on the large synthetic and real-world Persian text datasets that have been prepared. We also show that model with only trained on scene images are generalizable on unseen Persian document images without further fine-tuning or training. The proposed dataset, trained models, and the code will be publicly available at https://github.com/zobeirraisi/PTDR.