摘要
气象历史档案因年代久远、字迹模糊,现有的手写字符识别算法难以有效应对噪声,导致图像识别率低。针对此问题,应用集成学习思想,提出融合鲁棒非负矩阵分解和多尺度卷积的手写气象档案资料识别算法RNMF-MsCNN,以实现对噪声图像的准确识别。该模型分为重构器和分类器两个子模块,重构器模块扩展非负矩阵分解模型,学习低维潜在空间的特征,同时应用鲁棒的L21范数,解决噪声干扰等问题。分类器模块采用多尺度卷积加宽网络结构,更好地捕获节点特征。重构器和分类器相互依存、迭代更新,充分利用原图信息,消除噪声对图像的影响,更好地捕获手写数字的特征信息。通过迭代训练,模型的识别准确率达到了99%以上,优于其他对比算法。实验结果进一步验证了该模型在气象档案数据识别方面的有效性。
Due to the long age and fuzzy handwriting of meteorological historical archives,the existing handwritten character recognition algorithm is difficult to deal with the noise effectively,resulting in low image recognition rate.To solve this problem,a novel meteorological data handwritten recognition algorithm RNMF⁃MsCNN,which combines robust non⁃negative matrix factorization and multi⁃scale convolution,is proposed based on the idea of ensemble learning.The model is divided into two sub⁃modules:a reconfigurator and a classifier.The reconfigurator module extends the non⁃negative matrix factorization model and learns the features of low⁃dimensional latent space.At the same time,the robust L21 norm is applied to solve the problem of noise interference.The classifier module adopts multi⁃scale convolutional broadening network structure to capture node features better.Reconstructors and classifiers are interdependent and iteratively updated.Make full use of the original image information,eliminate the influence of noise on the image,and better capture the feature information of handwritten digits.Through iterative training,the recognition accuracy of the model reaches more than 99%,which is better than other comparison algorithms.The experimental results further validate the validity of the model in the identification of meteorological profile data.
作者
李子祺
勾志竟
雷鸣
LI Ziqi;GOU Zhijing;LEI Ming(Tianjin Meteorological Information Center,Tianjin 300074,China)
出处
《电子设计工程》
2025年第15期1-8,共8页
Electronic Design Engineering
基金
国家自然科学基金资助项目(41575156)
天津市气象局科研资助项目(202409ybxm08)。
关键词
深度学习
非负矩阵分解
多尺度卷积
气象档案
字符识别
deep learning
Non⁃negative Matrix Factorization
multi⁃scale convolutional
meteorological profile
character recognition
