基于注意力机制的海关报表识别方法研究  

Research on the method of customs statement identification based on attention mechanism

作　　者：万燕[1] 李毅凡 姚砺[1] WAN Yan;LI Yifan;YAO Li(College of Computer Science and Technology,Donghua University,Shanghai 201620,China)

机构地区：[1]东华大学计算机科学与技术学院,上海201620

出　　处：《智能计算机与应用》2024年第4期26-33,共8页Intelligent Computer and Applications

摘　　要：海关报表作为进出口业务中的重要材料,需要快速识别文本并录入系统,提高业务效率。海关报表图像通常存在字迹模糊粘连、字号过小和噪声污染等问题,增加了报表文本识别的难度。本文针对海关报表图片识别准确率低的问题,提出了基于注意力机制的海关报表识别方法。在DBNet模型中引入了注意力机制,提升小字符文本检测能力,使网络更加关注字符相关区域;在视觉模型中引入可变形卷积模块,扩大感受野,并将视觉特征和语义特征增强后通过门控机制实现多模态融合,提升对低质量字符的识别精度。实验结果表明,本文方法在海关报表低质量图像的检测和识别准确率方面领先其他方法。Customs statements,as important materials in import and export business,need to quickly identify the text and enter it into the system to improve business efficiency.Customs statement images usually have characteristics of blurred and sticky handwriting,too small font size and noise pollution,which increase the difficulty of statement text recognition.In this paper,we propose a customs statement recognition method based on the attention mechanism for the problem of low accuracy of customs statement image recognition.Attention mechanisms are introduced in DBNet model to enhance the small character text detection ability and make the network pay more attention to the character-related regions.The deformable convolution module is introduced in the visual model to expand the perceptual field,and the visual features and semantic features are enhanced to achieve multimodal fusion through the gating mechanism to improve the recognition accuracy of low-quality characters.The experimental results prove that this paper leads other methods in the detection and recognition accuracy of low-quality images in customs statements.

关 键 词：海关报表识别 注意力机制 文本识别 

分 类 号：TP391.43[自动化与计算机技术—计算机应用技术]