手写数字深度特征学习与识别

发布时间：2018-03-28 00:41

本文选题：深度学习　切入点：特征融合　出处：《计算机技术与发展》2016年07期

【摘要】：深度学习中的网络结构设计、特征提取与融合是数据挖掘和模式识别理论和行业应用中的关键问题。文中以相关领域中的典型应用问题手写数字识别和权威数据库MNIST为实验平台(包含七万个手写数字图像),探索了深度学习网络结构的设计和特征融合问题,保证研究结果的实用性、代表性和可参考性。所给方案的步骤是:首先,设计非监督深度学习网络,进行非监督高层语义特征学习,提取深度特征(DF),探索特征的高层认知特点;其次,对手写数字数据库进行非监督多特征提取,包括HOG(梯度方向直方图)特征、PCA(主成分分析)特征、LDA(判别分析)特征、像素分布特征、穿越次数特征和投影特征,构建手写数字典型特征库(Library of Typical Features,LTF);最后,构建深度有监督学习网络,有监督地融合深度特征DF和典型特征库。实验结果表明,相比于文献中的典型特征,该方案能够将手写数字识别的错误率有效降低50%。
[Abstract]:The design of network structure in deep learning, feature extraction and fusion is a key problem in data mining and pattern recognition theory and industry application. Typical application problems in relevant fields in the handwritten numeral recognition and authoritative database MNIST as the experimental platform (contains seventy thousand handwritten digital image), and explore the deep learning design and characteristics of the network the structure of the fusion problem, ensure the practicability of the research results, representative and reference. The plan steps are: firstly, supervision of deep learning network design, unsupervised learning high-level semantic feature extraction, feature depth (DF), characteristics of the high cognitive exploration characteristics; secondly, the handwritten digit database unsupervised feature extraction, including the HOG (gradient direction histogram), PCA (principal component analysis), LDA (discriminant analysis) feature pixel distribution, the number of crossing feature and projection feature Construction of syndrome, typical characteristics of handwritten digital library (Library of Typical Features, LTF); finally, the construction depth of supervised learning network, supervised feature fusion depth DF and typical feature database. The experimental results show that compared to the typical characteristics of the literature, this scheme can be hand written numeral recognition error rate decrease 50%.

【作者单位】：华南师范大学数学科学学院;
【基金】：国家自然科学基金资助项目(61273248;61075033)
【分类号】：TP391.41

【相似文献】