改进SLANet的OCR表结构识别方法  

Recognition Method of Improved OCR Table Structure for SLANet

作　　者：曹茂俊[1] 李悦 CAO Maojun;LI Yue(School of Computer and Information Technology,Northeast Petroleum University,Daqing 163318,China)

机构地区：[1]东北石油大学计算机与信息技术学院,黑龙江大庆163318

出　　处：《吉林大学学报(信息科学版)》2025年第1期98-106,共9页Journal of Jilin University(Information Science Edition)

基　　金：黑龙江省自然科学基金资助项目(LH2019F004);中石油科技术开发基金资助项目(2021DJ4001)。

摘　　要：针对传统的识别表结构方法难以充分学习多行多列合并、空白、嵌套单元格等复杂表结构以及提取特征过程中容易出现信息缺失的问题,提出了一种改进SLANet(Structure Location Alignment Network)的OCR(Optical Character Recognition)表结构识别方法。首先,利用轻量级CPU(Central Processing Unit)卷积神经网络并引入注意力机制,增强网络泛化和解释能力,将训练得到信息向量输入轻量级高低层特征融合模块中提取特征,并将输出特征通过特征解码模块对齐结构与位置信息,得到预测标签。实验表明,与EDD(Encoder-Dual-Decoder)、 TableMaster等模型相比,该方法准确率有显著提升,达到76.95%,TEDS(Tree-Edit-Distance-based Similarity)达到95.57%,显著增强了模型识别非常规复杂表结构能力,为识别表结构提供了一种优化策略。Traditional methods for identifying table structures are difficult to fully learn complex table structures such as merge cells with multiple rows and columns,blank cells,nested cells,and are lack of information in the process of extracting features.An OCR(Optical Character Recognition) table structure identification method based on improved SLANet(Structure Location Alignment Network) is proposed.Firstly,the lightweight CPU(Central Processing Unit) convolutional neural network is used and attention mechanism is introduced to enhance the generalization ability and explanation ability of the network.The information vector obtained by training is inputed into the lightweight high-low level feature fusion module to extract features,and then the outputted features are aligned with the structure and position information through the feature decoded module to obtain the prediction label.Experiments show that compared to EDD(Encoder-Dual-Decoder),TableMaster and other models,the accuracy of the proposed method has been significantly improved,reaching 76.95%,and the TEDS(Tree-Edit-Distance-based Similarity) has reached 95.57%,which significantly enhances the model's ability to identify complex table structures and provides an optimization strategy for identifying table structures.

关 键 词：识别表结构 结构位置对齐网络 注意力机制 基于树编辑距离的相似度 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]