融合深度置信网络的串联隐马尔科夫模型及其在脱机手写识别中的应用（英文）

发布时间：2018-03-04 07:35

本文选题：手写识别　切入点：隐马尔科夫模型　出处：《Frontiers of Information Technology & Electronic Engineering》2017年07期 　论文类型：期刊论文

【摘要】：在文档分析和模式识别领域,自由书写的脱机手写识别是一个非常具有挑战性的研究课题。近年来,为了充分探索隐藏在文档图像中的监督信息,许多研究工作试图将多层感知机以一种混合或串联的形式嵌入隐马尔科夫模型当中。然而,因为多层感知机学习能力的不足,学习到的特征对于后续的识别任务不一定是最优的。在本文中,我们针对自由书写的脱机手写识别提出一种基于深度结构的串联方法。在提出的模型中,深度置信网络被用于学习序列数据的紧致表示,隐马尔科夫模型被用于(子-)词的识别。我们在两个公开的数据集上验证了所提出的模型,这两个数据集是分别基于拉丁和阿拉伯语的RIMES和IFN/ENIT;我们还在Devanagari数据集上验证了所提出的模型,这个数据集是基于印度语的。大量的实验展示了所提出模型的优势,特别是相对于多层感知机-隐马尔科夫模型的串联方法。
[Abstract]:In the field of document analysis and pattern recognition, free writing offline handwriting recognition is a very challenging research topic. In recent years, in order to fully explore the hidden information in document images, Many researches have attempted to embed multilayer perceptron into Hidden Markov Model in a hybrid or series form. However, because of the lack of learning ability of multilayer perceptron, The features learned are not necessarily optimal for subsequent recognition tasks. In this paper, we propose a series method based on depth structure for free writing offline handwriting recognition. The deep confidence network is used to learn the compact representation of the sequence data, and the hidden Markov model is used to recognize the (sub-) words. We validate the proposed model on two open datasets. The two datasets are based on RIMES and IFN / ENITs in Latin and Arabic, respectively; we have also validated the proposed model on the Devanagari dataset, which is based on Hindi. A number of experiments have shown the advantages of the proposed model. In particular, the series method is relative to the multilayer perceptron-hidden Markov model.
【作者单位】： Department
【基金】：supported by the National Natural Science Foundation of China(No.61403353)
【分类号】：TP391.41

【相似文献】