基于改进CTPN网络的电气铭牌文本区域检测  

Detection of electric nameplate text area based on improved CTPN algorithm

作　　者：石煌雄 胡洋 蒋作[2] 潘文林[2] 杨凡 SHI Huang-xiong;HU Yang;JIANG Zuo;PAN Wen-lin;YANG Fan(School of Electrical and Information Technology,Yunnan Minzu University,Kunming 650500,Yunnan;School of Mathematics and Computer Science,YunnanMinzu University,Kunming 650500)

机构地区：[1]云南民族大学电气信息工程学院,云南昆明650500 [2]云南民族大学数学与计算机科学学院,云南昆明650500

出　　处：《云南民族大学学报（自然科学版）》2023年第1期96-102,共7页Journal of Yunnan Minzu University:Natural Sciences Edition

基　　金：国家自然科学基金(61866040,41961053);云南省教育厅科学研究基金(2020Y0242,2020Y0237).

摘　　要：针对现有的文本区域检测网络对图像特征提取不充分,导致文本行间距过小或字间距过大的电气铭牌文本区域检测准确率低、误检率高,提出一种融合ResNet的改进CTPN算法用于电气铭牌文本区域检测.该模型在卷积神经网络模块中引入残差连接和并行卷积核操作,提升模型对铭牌图像提取特征的能力.在长短期记忆网络模块采用双向门控单元,降低模型训练难度.经实验验证,改进后的CTPN模型F值(F-measure)明显提升,对铭牌文本区域误检率显著降低.For the existing text area detection network,the image feature extraction is insufficient,resulting in low accuracy and high false detection rate of the electric nameplate text area detection with too small text line spacing or too large word spacing.This paper proposes an improved CTPN algorithm fused with ResNetto detect the electric nameplate text area.This model introduces residual connection and parallel convolution kernel operations in the convolutional neural network module to improve the model s ability to extract features from the nameplate image.The two-way gating unit is used in the long-term and short-term memory network module to reduce the difficulty of model training.After experimental verification,the improved CTPN model F value(F-measure)is significantly improved,and the false detection rate of the nameplate text area is significantly reduced.

关 键 词：电气铭牌 文本区域检测 CTPN算法 卷积神经网络 长短期记忆网络 

分 类 号：TP391.1[自动化与计算机技术—计算机应用技术]