无人机输电线路巡检照片号牌文字识别方法  

Text Recognition Method for Power Pole Tower Number Plates in Unmanned Aerial Vehicle Inspection Photos Based on CTPN Algorithm

作　　者：李有春 汤春俊 梁加凯 林龙旭 徐敏 谢敏 LI Youchun;TANG Chunjun;LIANG Jiakai;LIN Longxu;XU Min;XIE Min(State Grid Zhejiang Electric Power Co.,Ltd.,Jinhua Power Supply Company,Jinhua 321017,China;School of Mechanical and Electrical Engineering,China Jiliang University,Hangzhou 310018,China;Jinhua Bada Group Co.,Ltd.,Jinhua 321001,China)

机构地区：[1]国网浙江省电力有限公司金华供电公司,浙江金华321017 [2]中国计量大学机电工程学院,浙江杭州310018 [3]金华八达集团有限公司,浙江金华321001

出　　处：《无线电工程》2024年第6期1560-1568,共9页Radio Engineering

基　　金：金华八达集团有限公司科技项目(BD2022JH-KXXM007)。

摘　　要：针对无人机巡检拍摄的高像素电力杆塔照片中杆塔号牌文字识别成功率低的问题,提出了一种改进连接文本区域网络(Connectionist Text Proposal Network,CTPN)算法。利用二维重叠滑动切割方法对输入图像进行切割,将主干网络Vgg16改为MobilenetV2对切割后图片进行卷积处理,同时在其中加入深度适配网络(Deep Adaptation Network,DAN)的注意力机制得到特征图;将卷积得到的特征图转化成序列输入至双向长短期记忆神经(Bi-directional Long Short-Term Memory,Bi-LSTM)网络学习序列特征,并通过全连接层得到建议框;加入重映射方法将建议框映射回原图,筛选整合映射到原图的建议框后,得到号牌文本框。将得到的文本框内的图像截取输入到卷积循环神经网络(Convolutional Recurrent Neural Network,CRNN)进行文字识别。实验结果表明,当切割框为456 pixel×256 pixel、横向重叠率为9%、纵向重叠率为8%时,识别精度可以达到87%。To address the problem of low recognition accuracy of tower number plate text in high-resolution power pole tower photos taken in drone inspection,an improved Connectionist Text Proposal Network(CTPN)algorithm is proposed.First,the input image is cut using a two-dimensional overlapping sliding cut method,and the backbone network Vgg16 is changed to MobilenetV2 to perform convolution on the cut images,while the attention mechanism of the Deep Adaptation Network(DAN)is added to obtain the feature maps.Second,the feature maps obtained by convolution are converted into sequences and input to the Bi-directional Long Short-Term Memory(Bi-LSTM)network to learn sequence features,and proposal boxes are obtained through the fully connected layer.Finally,the remapping method is added to map the proposal boxes back to the original image,and after filtering and integrating the proposal boxes mapped to the original image,the text box of the number plate is obtained.The image in the obtained text box is cropped and input to the Convolutional Recurrent Neural Network(CRNN)for text recognition.The experimental results show that when the cutting box is 456 pixel×256 pixel,the horizontal overlap rate is 9%,the vertical overlap rate is 8%,and the recognition accuracy can be up to 87%.

关 键 词：深度学习 高像素 场景文字识别 小目标 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]