Computer Engineering and Applications ›› 2021, Vol. 57 ›› Issue (24): 185-191.DOI: 10.3778/j.issn.1002-8331.2007-0098

• Pattern Recognition and Artificial Intelligence • Previous Articles     Next Articles

Text Detection in Natural Scenes Embedded Attention Mechanism

YANG Siqi, YI Yaohua, TANG Ziwei, WANG Xinyu   

  1. Department of Printing and Packaging, Wuhan University, Wuhan 430079, China
  • Online:2021-12-15 Published:2021-12-13

嵌入注意力机制的自然场景文本检测方法

杨锶齐,易尧华,汤梓伟,王新宇   

  1. 武汉大学 印刷与包装系,武汉 430079

Abstract:

For missed text detection and detected text deficiency, a text detection method embedded attention mechanism is proposed. Faster-RCNN and Feature Pyramid Network(FPN) are used as the basic framework. Embedded attention mechanism and improved anchor setting, which are designed by text characteristics, are utilized for more accurate text candidate regions. The scope of loss function is reset. The experiment results show that this method can significantly ensure the integrity of detected text information effectively and improve the recall rate and accuracy rate compared with existing methods, which can be exploited for text detection in natural scenes.

Key words: natural scene text detection, feature pyramid network, region proposal network, attention mechanism

摘要:

针对自然场景文本检测中存在的文本检测信息缺失、漏检的问题,提出了嵌入注意力机制的自然场景文本检测方法。利用Faster-RCNN目标检测网络和特征金字塔网络(FPN)作为基本框架;在区域建议网络(RPN)中嵌入注意力机制并依据文本的特点改进锚点(anchor)的设置,精确了文本候选区域;重新设定损失函数的作用范围。实验结果表明,该方法有效地保证文本检测信息的完整性,较之现有方法明显地提高了文本检测的召回率和准确率,能够应用于文本检测的实际任务中。

关键词: 自然场景文本检测, 特征金字塔网络, 区域建议网络, 注意力机制