Scholarlink

<p>Scene Text Recognition (STR) is an essential problem in computer vision, but a scarcity of real labeled data has slowed progress. Conventional dependence on artificial datasets typically fails in complex real-world scenarios because it cannot capture the richness of handwritten, artistic, and multilingual text. We present a novel hybrid architecture that addresses these challenges by merging Enhanced ResNet’s powerful spatial feature extraction with Gated Recurrent Convolutional Neural Networks’ (GRCNN) sequential modeling capabilities. ResNet enhances the extraction of discriminative visual qualities, whereas GRCNN employs Gated Recurrent Convolution Layers (GRCL) to capture contextual connections in text sequences. This technique strikes a reasonable balance between generalization and adaptation by including both synthetic and recently made available large-scale real-world datasets. Experimental results show that the proposed hybrid model, trained exclusively on synthetic datasets, achieves an average accuracy of 88.58% across various benchmarks. Incorporating actual datasets further improves this to 94.35%, surpassing several state-of-the-art approaches. The collaboration between architectural innovation and diverse training data pushes the boundaries of modern STR systems, providing a dependable and scalable solution that significantly improves identification accuracy in a wide range of circumstances.</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

Integrating CNN and BiLSTM for enhanced scene text recognition

  • 期刊论文

  • 发表日期：

    2026年3月27日

查看全文
  • Alshefaa Khattab,
  • Marwa Elpeltagy,
  • Farida Youness,
  • Ahmed Elshafei,
  • Ahmed Y. Khedr

摘要

Scene Text Recognition (STR) is an essential problem in computer vision, but a scarcity of real labeled data has slowed progress. Conventional dependence on artificial datasets typically fails in complex real-world scenarios because it cannot capture the richness of handwritten, artistic, and multilingual text. We present a novel hybrid architecture that addresses these challenges by merging Enhanced ResNet’s powerful spatial feature extraction with Gated Recurrent Convolutional Neural Networks’ (GRCNN) sequential modeling capabilities. ResNet enhances the extraction of discriminative visual qualities, whereas GRCNN employs Gated Recurrent Convolution Layers (GRCL) to capture contextual connections in text sequences. This technique strikes a reasonable balance between generalization and adaptation by including both synthetic and recently made available large-scale real-world datasets. Experimental results show that the proposed hybrid model, trained exclusively on synthetic datasets, achieves an average accuracy of 88.58% across various benchmarks. Incorporating actual datasets further improves this to 94.35%, surpassing several state-of-the-art approaches. The collaboration between architectural innovation and diverse training data pushes the boundaries of modern STR systems, providing a dependable and scalable solution that significantly improves identification accuracy in a wide range of circumstances.