改进的DenseNet的密集场景文本检测方法  

Improved DenseNet Text Detection Method for Dense Scenes

作　　者：吕鹏鹏 於跃成 齐秀芳 LYU Pengpeng;YU Yuecheng;QI Xiufang(School of Computer Science,Jiangsu University of Science and Technology,Zhenjiang 212100)

机构地区：[1]江苏科技大学计算机学院,镇江212100

出　　处：《计算机与数字工程》2025年第1期196-201,共6页Computer & Digital Engineering

摘　　要：场景文本检测是近年来极具挑战性的任务,针对自然场景中密集文本检测受限、漏检的特点,提出了面向场景文本的检测方法。首先,采用可变形ROI池代替平均池化层改进DenseNet网络作为特征提取网络,实现对不同尺度的文本进行自适应局部定位,然后通过卷积注意模块对多层级图像特征加权,增强文本特征。此外,在特征融合时引入可变卷积代替普通卷积,增加调整卷积核的方向向量,从而使采样网格自由变形,促使卷积核的形态更贴近文本形状。最后,在输出层引入辅助的双向门控循环单元,聚集文本区域。该模型相较于已有的方法在ICDAR2013数据集上提高了近1.11%,在ICDAR2015数据集上提高了近1.17%,一定程度上提高了检测的精确率。Scene text detection is a challenging task in recent years.Aiming at the limited and missing features of dense text detection in natural scenes,a scene image oriented detection method is proposed.Firstly,the deformable ROI pool is used to re⁃place the average pooling layer,and the improved denseNet network is used as the feature extraction network to realize the adaptive local localization of different scale texts.Then,the multi-level image features are weighted by convolution attention module to en⁃hance the text features.In addition,the variable convolution is introduced to replace the ordinary convolution in feature fusion,so that the direction vector of the convolution kernel is added to adjust,and the sampling grid is free to deform,so that the shape of the convolution kernel is closer to the shape of the text.Finally,an auxiliary bidirectional gating loop unit is introduced in the output lay⁃er to gather the text area.Compared with the existing methods,the model improves nearly 1.11%on ICDAR2013 data set and nearly 1.17%on ICDAR2015 data set,and improves the detection accuracy to a certain extent.

关 键 词：图像处理 文本检测 可变卷积 自然场景 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]