Network structure design,feature extraction and fusion in deep learning are key problems in data mining and pattern recognition theory and industry application. The design of deep learning network's structure and the problem of feature fusion is explored,taking handwritten numeral recognition and authoritative database MNIST,with 70 thousands of handwritten image,as the experiment platform, which guarantees the practicability,representation and reference of the research results. The solution step has been given. Firstly,the unsu-pervised deep learning network is designed,learning unsupervised high-level semantic features,extraction of depth features,and explora-tion of higher cognitive characteristics of features. Secondly,unsupervised features of handwritten database are extracted,including HOG, PCA,LDA and so on,construction of LTF. Finally,deep supervised learning network is built,fusion of deep features and the library of typical features with supervision. The result shows that this scheme can lower error rate of handwritten recognition by 50%,compared with the typical features of the present.%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%。
展开▼
