低质量海关报表字符识别模型研究

Research on character recognition model of low quality customs statements

海关报表单据图像质量差,其中字符往往有模糊、笔画缺失、笔画粘连和噪声污染等特点。本文针对海关报表单据中低质量字符识别准确率低的问题,提出了Enhanced-DBNet文本检测模型并改进ABINet文本识别模型。基于DBNet模型重新设计其主干网络,引入可变形卷积模块(DCN)扩大感受野,提高长文本识别能力;采用双向特征金字塔增强模块(FPEM),使网络具有更强的表征能力;引入特征融合模块(FFM)将图像高层次语义特征和低层次位置特征充分融合。针对形近字符难区分的问题,在ABINet模型中引入可变形注意力模块,使注意力集中在字符相关区域,捕获到更多的字符特征。对比实验结果表明,本文的模型在海关报表低质量字符上的检测和识别准确率优于当前其他模型。

The image quality of customs report documents is poor,and the characters are often characterized by blur,missing strokes,adhesion of strokes,and noise pollution.This paper proposes the Enhanced-DBNet text detection model and improves the ABINet text recognition model to solve the problem of low accuracy of low-quality character recognition in customs report documents.This paper redesigns the backbone network based on the DBNet model,introduces the deformable convolution module(DCN)to expand the receptive field and improves long text recognition capabilities,and uses the bidirectional feature pyramid enhancement module(FPEM)to make the network stronger representation capabilities and introduce feature fusion.The module(FFM)fully integrates high-level semantic features and low-level positional features of the image.To solve the problem of difficulty in distinguishing characters with similar shapes,a deformable attention module is introduced in the ABINet model to focus attention on character-related areas and capture more character features.Through comparative experiments,the model in this article has better detection and recognition accuracy on low-quality characters in customs reports than other current models.