基于高斯密度图的自然场景中文文字检测

Chinese text detection in natural scenes based on Gaussian density map

自然场景下中文文字检测任务字符面积较小且文字背景复杂,为此,该文提出了一种基于高斯密度图作为分割区域标签的文字检测改进方法。将文本区域标注从矩形框改变为高斯分布区域标签,在特征融合部分引入了交叉通道融合注意力机制,以提高网络的收敛性能,提出像素值一维投影法解决了文字密集区域输出的高斯密度图在不同标签区域出现重叠的问题。经过实验验证,该文方法对中文文字检测有较好的结果,证明了该方法的有效性。

The Chinese text detection task in natural scenes has small character area and complex text background,this paper proposes an improved method for text detection via Gaussian density map as segmented region labels.The text region label is changed from rectangular box to Gaussian distribution region label,channel⁃wise cross fusion transformer is introduced in the feature fusion part to improve the convergence performance of the network,the pixel⁃valued one⁃dimensional projection method is proposed to solve the problem of overlapping different labeled regions of Gaussian density maps output in dense regions of text.Experimental results show that the proposed method is effective for Chinese text detection.