%0 Journal Article
%T 融合空间特征的债券图表数据文本检测方法研究<br>Text Detection Method for Bond Chart Data Fusing Spatial Features
%A 李桂钢
%A 胡金蓉
%A 帅梓涵
%A 郎子鑫
%A 罗月梅
%J Hans Journal of Data Mining
%P 143-153
%@ 2163-1468
%D 2023
%I Hans Publishing
%R 10.12677/HJDM.2023.132014
%X 随着国家明确了金融业发展和改革的重点方向，我国金融数据信息化有了显著的发展和进步。基于债券图表数据的特定情况，人工处理债券图表数据存在效率低、成本高、安全性低等问题，用人工智能的方法来检测债券图表数据逐渐成为了当下的热门研究方向。由于债券图表数据在长时间存放、人为损坏等主客观因素下，会存在模糊、被污染等特点。对此本文使用了Swin-Transformer作为主干网络，它的特征提取能力较CNN (卷积神经网络)更为强大。并对模糊、污染的区域设计了方向感知模块，使其对文本区域的识别正确率更高。实验结果表明，该网络比其它文本检测算法在准确率、召回率、F1值上都有明显提升。<br />
With the clear direction of financial development and reform, China’s financial data information technology has made significant progress and development. Based on the specific situation of bond chart data, manual processing of bond chart data is inefficient, high cost, low security, and so on. Using artificial intelligence to detect bond chart data has gradually become a hot research direction. Because the bond chart data is stored for a long time and damaged artificially, it will be fuzzy and polluted. In this paper, Swin-Transformer is used as the backbone network, and its feature extraction ability is stronger than that of CNN (convolution network). Direction perception module is de-signed for blurred and contaminated areas, which makes the recognition of text areas more accurate. The experimental results show that the network improves the accuracy, recall and F1 value significantly compared with other text detection algorithms.
%K 债券图表数据，文本检测，Swin-Transformer，方向感知模块，Bond Chart Data
%K Text Detection
%K Swin-Transformer
%K Direction-Aware Module
%U http://www.hanspub.org/journal/PaperInformation.aspx?PaperID=64198