摘要
针对目前主流的基于分割的文字检测方法中由于需要复杂的后处理过程保证检测精度,通常难以实现高检测速度的问题,提出一种应用位置注意力模块和金字塔注意力网络2种注意力机制的方法.首先用金字塔注意力网络对图像进行特征提取及语义分割;同时将位置注意力模块应用于高层特征,通过加强图像中相似物体的权重加强对文字的检测效果;最后进行简单有效的后处理,在实现较高检测准确度的前提下提高检测速度.实验结果表明,在Total-text数据集中,采用更轻量化的骨干网络时,所提方法在检测速度上优势明显;采用更深层的骨干网络时,所提方法的检测准确度领先2.0%.
In view of the issue that current mainstream segmentation-based text detection methods is difficult to achieve high detection speed due to complex post-processing to ensure detection accuracy,a scene text detec-tion method is proposed which applies pyramid attention network and position attention module.First,it adopts pyramid attention network to perform feature extraction and semantic segmentation.Meanwhile,it adopts po-sition attention module in high-level features,which strengthens the weights of similar objects in the image to enhance the effect of text detection.Finally,it adopts a simple and effective post-processing algorithm to in-crease detection speed under the premise of high detection accuracy.Experimental results show that in To-tal-text datasets,using light-weight backbone network,the method has great advantages on detection speed,and while using deeper backbone network,the method achieves the state of the art result and has a 2.0%lead on detection accuracy.
作者
王延昭
顾晓东
Wang Yanzhao;Gu Xiaodong(Department of Electronic Engineering,School of Information Science and Technology,Fudan University,Shanghai 200438)
出处
《计算机辅助设计与图形学学报》
EI
CSCD
北大核心
2021年第12期1908-1915,共8页
Journal of Computer-Aided Design & Computer Graphics
基金
国家自然科学基金面上项目(61771145).
