摘要
针对深度学习中对任意形状文本检测准确率不高的问题,提出了一种结合特征金字塔网络(feature pyramid network,FPN)和内核尺度扩展算法的文本检测网络模型。特征金字塔网络能够提取卷积层中更加鲁棒的特征,融合后生成不同尺度的特征内核;内核尺度扩展算法将生成的最小特征内核逐渐扩展为包围完整文本实例的特征图。同时为了针对自然场景中难以检测的文本实例,在训练阶段加入了在线难例挖掘(online hard example mining,OHEM)的方法,并以迁移学习的方式采用2种不同训练策略进行训练。仿真结果表明,该算法模型在不同数据集上具有良好的检测性能。
Aiming at the problem of low accuracy in detecting arbitrary shape text in deep learning,a text detection method combining feature pyramid network(FPN)and kernel scale extension algorithm is proposed in this paper.The feature pyramid network can extract more robust features in the convolutional layer and generate different scale feature kernels after fusion.The kernel scale expansion algorithm gradually expands the generated minimum feature kernel into a feature map which surrounds the entire text instance.At the same time,for the text examples which are difficult to detect in natural scenes,the online hard example mining(OHEM)method was added during the training phase,and two training strategies were used to train in the form of transfer learning.Simulation results show that the algorithm model has good detection performance on different datasets.
作者
林金朝
文盼
庞宇
LIN Jinzhao;WEN Pan;PANG Yu(School of Communication and Information Engineering,Chongqing University of Posts and Telecommunications,Chongqing 400065,P.R.China;Chongqing Key Laboratory of Photoelectronic Information Sensing and Transmitting Technology,Chongqing 400065,P.R.China)
出处
《重庆邮电大学学报（自然科学版）》
CSCD
北大核心
2022年第1期155-163,共9页
Journal of Chongqing University of Posts and Telecommunications(Natural Science Edition)
基金
国家自然科学基金(61301124,61671091)。
关键词
深度学习
文本检测
特征金字塔
内核扩展
deep learning
text detection
feature pyramid
kernel expansion
