Scholarlink

<p>Scene text recognition is a hot topic in the field of computer vision. Scene text recognition technology can enable computers to extract text information quickly and accurately in complex scenes. At present, the Attention-based Encoder-Decoder method has become the mainstream of scene text recognition, but it often suffers from the problem of attention drift. In response to this problem, this paper proposes TRBA++ to alleviate the problem of attention drift by improving the attention method in the decoder. Specifically, by calculating the relationship between feature sequences with different distances, the feature sequences with farther distances can be connected, and more visual information can be obtained so that the information from the visual and semantic can be better combined. Through comparative experiments, it can be seen that the scene text recognition method TRBA++ proposed in this paper performs well on multiple public datasets, especially on SVT and IC13 857 in the regular dataset, with the accuracy increasing by 1.4% and 2.1% respectively, and on SVTP and CT in irregular datasets increases the accuracy by 3.1% and 2.4%, respectively.</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

Feature enhanced attention decoder for scene text recognition

  • 期刊论文

  • 发表日期：

    2025年4月10日

查看全文
  • Miaomiao Xu,
  • Jiang Zhang,
  • Lianghui Xu,
  • Wushour Silamu,
  • Yanbing Li

摘要

Scene text recognition is a hot topic in the field of computer vision. Scene text recognition technology can enable computers to extract text information quickly and accurately in complex scenes. At present, the Attention-based Encoder-Decoder method has become the mainstream of scene text recognition, but it often suffers from the problem of attention drift. In response to this problem, this paper proposes TRBA++ to alleviate the problem of attention drift by improving the attention method in the decoder. Specifically, by calculating the relationship between feature sequences with different distances, the feature sequences with farther distances can be connected, and more visual information can be obtained so that the information from the visual and semantic can be better combined. Through comparative experiments, it can be seen that the scene text recognition method TRBA++ proposed in this paper performs well on multiple public datasets, especially on SVT and IC13 857 in the regular dataset, with the accuracy increasing by 1.4% and 2.1% respectively, and on SVTP and CT in irregular datasets increases the accuracy by 3.1% and 2.4%, respectively.