Scholarlink

This chapter further explores how to model long-term contextual dependencies in text images. A dynamic temporal residual learning mechanism is proposed to model the contextual information in feature sequences by introducing the residual learning method into the temporal dimension of an RNN encoder. An attention rectification method is also proposed to mitigate the attention drift problem in the decoder.

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

Dynamic Temporal Residual Learning and Attention Rectification

  • 章节

  • 发表日期：

    2025年

查看全文
  • Liangrui Peng,
  • Ruijie Yan

摘要

This chapter further explores how to model long-term contextual dependencies in text images. A dynamic temporal residual learning mechanism is proposed to model the contextual information in feature sequences by introducing the residual learning method into the temporal dimension of an RNN encoder. An attention rectification method is also proposed to mitigate the attention drift problem in the decoder.