基于深度学习手写字符的特征抽取方法研究.pdfVIP

2017 年 软 件 2017, Vol. 38, No. 1 第 38 卷 第 1 期 COMPUTER ENGINEERING SOFTWARE 国际 IT 传媒品牌 基于深度学习手写字符的特征抽取方法研究 邹 煜，刘兴旺 ( 中南民族大学 计算机科学学院，武汉 430074) 摘 要: 当前对深度学习单层训练算法的研究工作较少，本文采用数据量、隐层节点和感受野大小，分析自动编 码器和 K-means 算法在训练深度网络抽取特征上的表现。发现自动编码器对数据量，隐层节点敏感，且学习率与数据 量和感受野呈负相关；数据量一定后，K-means 对数据量不敏感，且感受野大小的选取对该算法发挥性能至关重要。 在实际应用中给自动编码器加入稀疏性控制是必要的。实验结果表明，本文的研究工作，对用自动编码器或 K-means 训练深度网络有一定的参考借鉴意义。 关键词: 深度学习；K-means；自动编码器；感受野；数据量；隐层节点 中图分类号: TP391.43 文献标识码: A DOI ：10.3969/j.issn.1003-6970.2017.01.006 本文著录格式：邹 煜，刘兴旺. 基于深度学习手写字符的特征抽取方法研究[J]. 软件，2017 ，38 （1）：23-28 Research of Feature Extraction Method for Offline Handwritten Character Based on Deep Learning ZOU Yu, LIU Xing-wang (College of Computer Science, South-Central University for Nationalities, Wuhan 430074) 【Abstract 】: In currently, there are few studies of monolayer training algorithms about deep learning. In this paper, We analyze the performance of the AutoEncoder and K-means in training deep network and feature extraction on three points:data volume, hidden layer nodes and receptive field. We find that automatic encoder is sensitive to data volume and hidden layer nodes,and learning rate has a negative correlation with data volume and receptive field. After a certain amount of data, K-means is not sensitive to data volume, so select receptive field size plays a crucial performance for the algorithm. In practice, added sparsity control into AutoEncoder is necessary. Experimental results shows that studies of this paper has reference value for training deep network with AutoEncoder or K-means. 【K