基于高斯密度图估计的自然场景汉字检测

针对自然场景下中文小文本难以定位的问题，提出了基于高斯密度图估计的并行深度网络对自然场景汉字进行检测。首先将中文数据集中的汉字位置信息转换为高斯文字密度图；其次引入一种多级并行连接结构，提高网络细节信息捕捉能力；最后再融合网络中的上采样特征信息得到高精度文字密度图，最终实现对文字区域的定位。在中文数据集CTW(Chinese text in the wild)上进行了实验，实验结果表明提出方法准确率和召回率均有较大提升，证明了该方法的可行性和准确性。

Chinese character detection in natural scene based on Gaussian density map estimation

Aiming at the nodus of samll Chinese text detection in natural scene, this paper proposed a parallel deep network based on Gaussian density map estimation to detect Chinese characters in natural scene. Firstly, it converted the position information of Chinese characters into a Gaussian text density map. Secondly, in order to improve the ability to capture network details, it used a multi-level parallel connection structure. Ultimately, the network combined the upsampling operation to fuse the feature information in the network to obtain a high-precision text density map, then realized the positioning of the text area through post-processing. This paper experimented on Chinese dataset CTW. The results show that the precision and recall rates of the method are both improved, demonstrate the feasibility and accuracy of the method.