[1]张墨逸,袁小芳,张彬,等.基于PCA-LDA和RIC-CNN的旋转无关手写识别[J].计算机技术与发展,2025,(04):93-99.[doi:10.20165/j.cnki.ISSN1673-629X.2024.0384]
　ZHANG Mo-yi,YUAN Xiao-fang,ZHANG Bin,et al.Rotation-independent Handwriting Recognition Based on PCA-LDA and RIC-CNN[J].,2025,(04):93-99.[doi:10.20165/j.cnki.ISSN1673-629X.2024.0384]
点击复制

基于PCA-LDA和RIC-CNN的旋转无关手写识别()

《计算机技术与发展》[ISSN:1006-6977/CN:61-1281/TN]

卷:
期数:
2025年04期
页码:
93-99
栏目:
人工智能
出版日期:
2025-04-10

文章信息/Info

Title:
Rotation-independent Handwriting Recognition Based on PCA-LDA and RIC-CNN
文章编号:
1673-629X(2025)04-0093-07
作者:
张墨逸; 袁小芳; 张彬; 陈海燕
兰州理工大学 计算机与通信学院,甘肃 兰州 730050
Author(s):
ZHANG Mo-yi; YUAN Xiao-fang; ZHANG Bin; CHEN Hai-yan
School of Computer and Communication,Lanzhou University of Technology,Lanzhou 730050,China
关键词:
手写字符识别; 非分离旋转矩不变量; 旋转不变坐标卷积神经网络; 正交投影; 主成分分析; 线性判别分析; 全连接神经网络
Keywords:
handwriting character recognition; non-detached moment of rotation invariant; rotationally invariant coordinate convolutional neural network; orthographic projection; principal component analysis; linear discriminant analysis; fully connected neural networks
分类号:
TP391
DOI:
10.20165/j.cnki.ISSN1673-629X.2024.0384
摘要:
旋转无关手写字符识别对提高手写识别系统的实际应用准确性和鲁棒性至关重要。 针对旋转字符识别准确率低、旋转不变坐标卷积神经网络(RIC-CNN)模型耗时长的问题,提出一种新的旋转无关手写字符识别方法。 首先,在原 RIC-CNN 模型中引入非分离旋转矩不变量方法,并创建特征提取层,提取旋转不变数据特征,解决数据提取不充分的问题。 接着,应用基于正交投影的 PCA-LDA 算法,该算法通过引入随机矩阵对特征进行正交投影,并将主成分分析(PCA)和线性判别分析(LDA)进行结合构建协方差矩阵和散射矩阵,从特征中筛选出有效的旋转不变特征,提高识别精度。 最后,采用全连接神经网络(FCNN)对提取的特征进行分类识别。 实验在空中手写数据集、MNIST 数据集、中科院数据集与哈工大数据集上进行了验证。 结果表明,该方法显著提升了旋转手写字符的识别精度,其中数字旋转数据的准确率达到了 97. 72% 。 以 MNIST 数据集为例,该方法识别字符的时间仅为 14. 16 秒,相比单独使用 RIC-CNN 模型,时间成本减少了 19. 66 秒,充分证明了该方法的有效性。
Abstract:
Rotation-independent handwriting character recognition is essential to improve the accuracy and robustness of handwriting rec-ognition systems in practical applications. Aiming at the problems of low accuracy of rotated character recognition and long time consumption of the Rotation Invariant Coordinate Convolutional Neural Network ( RIC - CNN) model, a new rotation - independent handwritten character recognition method was proposed. Firstly,the non-detached rotational moment invariant method was introduced into the original RIC-CNN model,and a feature extraction layer was created to extract the rotational invariant data features to solve the problem of insufficient data extraction. Then, the PCA -LDA based on orthogonal projection was applied,which introduced random matrices to project the features orthogonally, and combined Principal Component Analysis ( PCA) and Linear Discriminant Analysis (LDA) to construct covariance matrix and scattering matrix,so as to screen out effective rotational invariant features from the features and improve the recognition accuracy. Finally,the Fully -Connected Neural Network ( FCNN) was used to classify and identify the extracted features. The experiments were verified on the aerial handwriting dataset,the MNIST dataset,the Chinese Academy of Sciences dataset and the Harbin Institute of Technology dataset. The results show that the proposed method significantly improves the recognition accuracy of rotated handwritten characters,and the accuracy of rotation data reaches 97. 72% . Taking the MNIST dataset as an example,the time for the proposed method to recognize characters is only 14. 16 seconds,which is 19. 66 seconds less than the RIC-CNN model alone,which fully proves the effectiveness of the method.
更新日期/Last Update: 2025-04-10