轻量级柱面线杆标识牌字符识别算法  

Character recognition algorithm on cylindrical utility pole nameplate based on lightweight network

作　　者：林绍福 李松静 刘希亮 LIN Shao-fu;LI Song-jing;LIU Xi-liang(College of Software Engineering,Faculty of Information Technology,Beijing University of Technology,Beijing 100124,China;Beijing Institute of Smart City,Faculty of Information Technology,Beijing University of Technology,Beijing 100124,China)

机构地区：[1]北京工业大学信息学部软件学院,北京100124 [2]北京工业大学信息学部北京智慧城市研究院,北京100124

出　　处：《计算机工程与设计》2023年第8期2498-2505,共8页Computer Engineering and Design

基　　金：集美大学航海学院-船舶辅助导航技术国家地方联合工程研究中心开放基金项目(JMCBZD202013)。

摘　　要：为全面、准确、快速地提取柱面电线杆标识牌信息,提出一种轻量级柱面电线杆标识牌字符识别算法Tiny-DBNet-CRNN。对柱面图像进行反投影矫正展平;融合注意力机制,利用深度可分离卷积残差块,构建轻量级文本检测网络分割出文本区域;构建字符识别模型CRNN输出标识牌字符信息。采用真实场景数据和ICDAR 2015数据进行实验,结果与当前流行模型相比,Tiny-DBNet-CRNN字符识别正确率提升了40.3%,达95.11%;在精度下降0.60%的微小损失下,检测速度提升3倍,参数规模上总体下降45.15%。To accurately,quickly and efficiently extract the information of cylindrical pole identification plate,a lightweight character recognition algorithm Tiny-DBnet-CRNN was proposed.The cylindrical image was corrected and flattened by back projection.A lightweight text detection network was constructed via attention mechanism and deep separable convolutional residual blocks to segment the texts in the nameplates.The character sequence recognition model CRNN was employed to deliver the character information of pole signs.The image data of real scene cylindrical utility pole nameplates and ICDAR2015 dataset were used for experiments.The results show that the accuracy of model recognition is increased by 40.3%,reaches 95.11%.Compared with original DBNet,its detection speed is improved by 3 times on the premise of the precision reduction of 0.60%,while the parameter size is decreased by 45.15%.

关 键 词：电线杆标识牌 反投影算法 轻量级网络 深度可分离卷积 注意力机制 DBNet算法 CRNN算法 

分 类 号：TP391.43[自动化与计算机技术—计算机应用技术]