面向结构化文本图像的四元数卷积神经网络模型设计

打开文本图片集

中图分类号：TP18 文献标识码：A 文章编号：2096-4706（2026）04-0116-06

Abstract：This paper designs a Quatermion Convolutional Neural Network （QCNN） for structured text images to address the problemof insufficient feature couplingandrobustnessdrop inreal-valued Convolutional Neural Networks caused by characteradhesionand color fraud incomplex color CAPTCHA recognition.This network encodes RGB pixels as a vector fieldusing purequaterions，ndachieves fullprocesshypercomplex processng throughHamilton product convolutionasor ReLU，and thegeneralizedHRdiferentialoptimizer.ItiscompardwithbaseliemodelssuchasResNet-18ona3self-built strong interference dataset. The experimental results show that the character level accuracy of QCNN reaches 97.8% ，and the sequence level accuracy is 96.4% ，significantly better than existing real and complex models，providing a new approach for high interference structured text image recognition.

KeyWords： Quaternion Convolutional Neural Network; structured text image; CAPTCHA recognition; color vecto1 modeling;Hamilton product

0 引言

验证码是一种区分用户是计算机还是人的公共全自动程序，是抵御网络爬虫、防止恶意注册和保护数据资源的第一道屏障[。（剩余7843字）