面向结构化文本图像的四元数卷积神经网络模型设计OA
Design of Quaternion Convolutional Neural Network Model for Structured Text Images
针对复杂彩色验证码识别中字符粘连与色彩欺诈所引起的实数卷积神经网络特征耦合不足与鲁棒性骤降的问题,文章设计了一种针对结构化文本图像的四元数卷积神经网络(QCNN).该网络将RGB像素用纯四元数表示并全部编码为矢量场,通过Hamilton积、Phasor ReLU及广义HR微分优化器全程实现超复数化,同时在本课题自建的3 000 张强干扰数据集上与ResNet-18 等基线模型展开对比.实验结果表明,QCNN的字符级准确率达到 97.8%,序列级准确率为96.4%,均优于现有的实数和复数模型,为高干扰结构化文本图像识别提供新的思路.
This paper designs a Quaternion Convolutional Neural Network(QCNN)for structured text images to address the problem of insufficient feature coupling and robustness drop in real-valued Convolutional Neural Networks caused by character adhesion and color fraud in complex color CAPTCHA recognition.This network encodes RGB pixels as a vector field using pure quaternions,and achieves full process hypercomplex processing through Hamilton product convolution,Phasor ReLU,and the generalized HR differential optimizer.It is compared with baseline models such as ResNet-18 on a 3 000 self-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.
马阳
江西工业职业技术学院,江西 南昌 330096||江西开放大学,江西 南昌 330046
信息技术与安全科学
四元数卷积神经网络结构化文本图像验证码识别色彩矢量建模Hamilton积
Quaternion Convolutional Neural Networkstructured text imageCAPTCHA recognitioncolor vector modelingHamilton product
《现代信息科技》 2026 (4)
116-121,6
江西省教育厅科学技术研究项目(GJJ2210602)
评论