摘要
由于自然场景中的图像背景复杂、文字排列不规则、光照条件不确定等因素,文字检测难度较大,且传统检测方法的效果并不理想。在研究文字分割检测方法PSENet(Progressive Scale Expansion Network)的基础上,提出了一种针对自然场景文字检测的改进方法。该方法由卷积神经网络提取特征模块,再通过渐进式规模扩张对文字区域进行分割检测。改进点主要是使用高精度的语义分割网络RefineNet(Refinement Network),对卷积网络特征提取模块进行优化,且增加较多的残差连接及链式池化,提高网络对文字区域的检测精度。通过对数据集ICDAR2015的实验结果对比,表明所提出的改进算法在精度方面略高于改进前,且能更好地解决文字粘连问题。
Due to the problematic scene background,irregular arrangement of text,and uncertain lighting conditions in natural scenes,text detection is difficult,and the traditional detection method is not ideal.In the study of the text segmentation detection method Progressive Scale Expansion Network(PSENet),an improved method for text detection in a natural scene is proposed.The improved model mainly uses the convolutional neural network to extract feature modules and performs segmentation detection on the text area through progressive scale expansion.The improvement points mainly uses a high-precision semantic segmentation network(RefineNet),optimizing the volume and network feature extraction modules,adding more residual connections and chain pooling,and improving the network's detection accuracy of the text area.Comparing the experimental results on the data set(ICDAR2015),the proposed improved algorithm is slightly more accurate than the previous algorithm and can better solve the problem of text conglutination.
作者
施漪涵
仝明磊
SHI Yihan;TONG Minglei(School of Electronics and Information Engineering,Shanghai University of Electric Power,Shanghai 200090,China)
出处
《上海电力大学学报》
CAS
2021年第1期73-77,共5页
Journal of Shanghai University of Electric Power
关键词
文字检测
图像分割
特征融合
text detection
image segmentation
feature fusion
作者简介
通信作者:仝明磊(1976—),男,博士,副教授。主要研究方向为利用人工智能技术的三维视觉重建(Visual Slam的核心技术)。E-mail:tongminglei@gmail.com。
