基于多级特征选择的自然场景文本识别算法  

Natural scene text recognition algorithm based on multilevel feature selection

作　　者：李利荣[1,2] 张开 张云良 乐玲 周蕾 巩朋成 LI Lirong;ZHANG Kai;ZHANG Yunliang;YUE Ling;ZHOU Lei;GONG Pengcheng(School of Electrical and Electronic Engineering,Hubei University of Technology,Wuhan,Hubei 430064,China;Hubei Engineering Research Center for New Energy and Power Grid Equipment Safety Monitoring,Wuhan,Hubei 430064,China)

机构地区：[1]湖北工业大学电气与电子工程学院,湖北武汉430064 [2]新能源及电网装备安全监测湖北省工程研究中心,湖北武汉430064

出　　处：《光电子．激光》2022年第5期479-487,共9页Journal of Optoelectronics·Laser

基　　金：国家自然科学基金(62071172);新能源及电网装备安全监测湖北省工程研究中心开放研究基金(HBSKF202121)资助项目。

摘　　要：针对现有场景文本识别方法只关注局部序列字符分类,而忽略了整个单词全局信息的问题,提出了一种多级特征选择的场景文本识别(multilevel feature selection scene text recognition,MFSSTR)算法。该算法使用堆叠块体系结构,利用多级特征选择模块在视觉特征中分别捕获上下文特征和语义特征。在字符预测过程中提出一种新颖的多级注意力选择解码器(multilevel attention selection decoder,MASD),将视觉特征、上下文特征和语义特征拼接成一个新的特征空间,通过自注意力机制将新的特征空间重新加权,在关注特征序列的内部联系的同时,选择更有价值的特征并参与解码预测,同时在训练过程中引入中间监督,逐渐细化文本预测。实验结果表明,本文算法在多个公共场景文本数据集上识别准确率能达到较高水平,特别是在不规则文本数据集SVTP上准确率能达到87.1%,相比于当前热门算法提升了约2%。Aiming at the problem that existing scene text recognition methods only focus on the classification of local sequence characters and ignore the global information of the entire word,a multilevel feature selection scene text recognition(MFSSTR) algorithm is proposed.The algorithm uses a stacked block architecture and applies a multilevel feature selection module to capture contextual and semantic features in visual features.In the process of character prediction,a novel multilevel attention selection decoder(MASD) is proposed,which combines visual features,context features and semantic features into a new feature space,and re-weights the new feature space through a self-attention mechanism.While paying attention to the internal relations of the feature sequence,select more valuable features and participate in decoding prediction.At the same time,intermediate supervision is introduced in the training process to gradually refine the text prediction.The experimental results show that the algorithm in this paper can reach a high level of recognition accuracy on multiple public scene text data sets.In particular,the accuracy rate can reach 87.1% on the irregular text data set SVTP,which is improved compared with the current popular algorithms by about 2%.

关 键 词：场景文本识别 特征序列 自注意力机制 多级注意力选择解码器 中间监督 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]