基于感受野增强和改进型损失函数的文本检测  

Text detection based on enhanced receptive fieldand improved loss function

作　　者：方承志 张子渊 李晨曦 FANG Cheng-zhi;ZHANG Zi-yuan;LI Chen-xi(College of Electronic and Optical Engineering、College of Microelectronics,Nanjing University of Posts and Telecommunications,Nanjing 210023,Jiangsu China)

机构地区：[1]南京邮电大学电子与光学工程学院、微电子学院,江苏南京210023

出　　处：《微电子学与计算机》2021年第4期11-16,共6页Microelectronics & Computer

基　　金：国家自然科学基金面上资助项目(61271334,61073115)。

摘　　要：针对目前主流文本检测算法对多方向文本检测效果显著,但对曲线文本检测不佳的问题,提出一种基于感受野增强和改进型损失函数的文本检测算法,适用于曲线及任意方向文本.该算法基于TextSnake网络改进,加入ASPP网络增大感受野以获得在深层网络中更高层次的语义信息,并逐层进行特征融合提高对复杂背景下的文本检测精度;同时针对数据正负例样本不平衡问题在损失函数部分采用加权的Focal Loss,提高检测过程中难检测像素点的损失.该算法在Total-Text和ICDAR2015数据集上测试,实验结果表明该算法在准确率和F值上均有较好的表现.The current mainstream text detection algorithms have significant effect on multi-directional text detection,but they are not good at detection of curve text.A text detection algorithm based on enhanced receptive field and improved loss function is proposed in this paper,which is also suitable for curve and arbitrary text.The algorithm is further improved based on TextSnake network.Improved ASPP network is added to increase receptive field to obtain higher level semantic information in the deep network,and feature fusion is performed layer by layer to improve the detection accuracy in the complex background.In the part of loss function,the weighted Focal Loss is used to improve the loss of pixels difficult in the detection process.The algorithm is tested on the Total-Text datesets and ICDAR2015 datesets,and the experimental results show that the algorithm performs well in both precision and F-score.

关 键 词：文本检测 感受野 ASPP网络 Focal Loss 正负例样本不平衡 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]