注意力监督策略下的自然场景文本检测算法

Text Detection Algorithm for Natural Scenes under Attention Supervision Strategy

已有的场景文本检测算法在处理任意形状的文本区域时,会将多余的背景区域也包含在内,为了自动学习聚焦任意形状的文本区域,提出一种注意力监督策略下的文本检测算法.首先使用深度残差网络作为骨架网络提取包含多尺度信息的特征图,然后通过注意力掩膜生成模块将融合特征图转换生成注意力掩膜,再通过背景抑制模块,利用注意力掩膜监督生成下一级特征图,最后经过一系列卷积操作生成分割掩膜,处理优化后得到最终的文本检测结果.实验表明,所提算法在ICDAR2015数据集上的多指标综合表现优越,其中F值相较对比算法提高了2.1%.

The existing scene text detection algorithms will include redundant background area when dealing with arbitrary shape text regions.In order to automatically learn to focus on arbitrary shape text regions,a text detection algorithm is proposed under attention supervision strategy.Firstly,deep residual network is used as the skeleton network to extract feature maps containing multi-scale information,and then the fused feature maps are converted to generate attention masks through the attention mask generation module,and through the background suppression module,the next level feature map is generated by attention mask.Fi-nally,the segmentation masks are generated through a series of convolution operations,and the final text detection results are obtained after post-processing optimization.Experimental results show that the pro-posed algorithm has superior multi-index comprehensive performance on the ICDAR2015 dataset,and the F-measure is 2.1%ahead of the comparison algorithm.