基于深度学习的手写汉字识别技术研究

发布时间：2018-04-16 10:30

本文选题：手写汉字识别 + 深度学习　；参考：《哈尔滨理工大学》2017年硕士论文

【摘要】：世界上使用人数最多、信息存储最简练、运用最广泛的语言文字,是中国传统文化和世界文化不可或缺的重要部分,所以针对汉字图像的识别研究具有非常重要的实用价值。手写汉字识别在日常的办公自动化、邮件分拣、金融等领域都具有非常很好的前景,手写汉字识别的研究可以使人们的日常生活更便捷。但是由于汉字的类别多,并且结构复杂,存在大量的相似汉字,再加上不同的人有不同的书写风格,在不同的书写环境、书写方式情况下,手写出来的汉字图像从种类和相似度等方面都很复杂,这就使得手写汉字识别在模式识别领域一直是难点同时也是研究热点。近年来,深度学习成为机器学习领域最热门的研究内容,尤其是在图像识别领域取得了非常好的效果。它能将复杂函数用更简洁的方式表达,可以自动获取样本概率分布同时在学习样本特征上具有优势。因此本文将深度学习模型应用到脱机手写汉字识别任务当中,通过使用深度学习中不同的方法和模型,进一步提高手写汉字识别的准确性。本文研究内容主要分以下几个方面:首先从深度信念网络入手,分析了传统手写汉字识别方法的不足,提出了一种基于深度信念网络融合模型对手写汉字识别的方法。该方法首先用修正的二次判别函数分类器识别较为简单的汉字,而较为复杂的汉字图像交给深度信念网络模型来处理,具体的分工过程通过定义可信度来协调两个分类器在识别任务中的分工,从而达到理想的识别效果。然后从卷积神经网络入手,卷积神经网络优于传统的图像分类方法在特征提取方面,所以将卷积神经网络应用在脱机手写汉字识别问题中。分析了卷积在手写汉字处理中的优缺点,针对相似手写汉字分类问题,进一步改进卷积神经网络,充分利用网络中特征提取的优势,二分类器则采用经典的SVM分类器,两者结合更好的对相似手写汉字识别分类。本文实验选用MNIST手写字、中国科学院自动化研究所采集的CASIA-HWDB1.1、北京邮电大学采集的HCL2000数据集。实验表明,本文提出的基于卷积神经网络的手写汉字识别方法能取得较好的识别效果。
[Abstract]:It is an indispensable part of Chinese traditional culture and world culture to use the largest number of people in the world, the most concise information storage, the most widely used language and characters, so the research on Chinese character image recognition has very important practical value.Handwritten Chinese character recognition has a very good prospect in the fields of office automation, mail sorting, finance and so on. The research of handwritten Chinese character recognition can make people's daily life more convenient.But because of the variety of Chinese characters, the complexity of their structure, the existence of a large number of similar characters, and the fact that different people have different writing styles, in different writing environments and ways of writing,Handwritten Chinese character images are very complex in terms of type and similarity, which makes handwritten Chinese character recognition difficult and research hotspot in the field of pattern recognition.In recent years, depth learning has become the most popular research content in machine learning field, especially in the field of image recognition.It can express the complex function in a more concise way, and it can automatically obtain the probability distribution of the sample and have the advantage in learning the feature of the sample.In this paper, the depth learning model is applied to the off-line handwritten Chinese character recognition task, and the accuracy of handwritten Chinese character recognition is further improved by using different methods and models in depth learning.The main contents of this paper are as follows: firstly, from the perspective of depth belief network, the shortcomings of traditional handwritten Chinese character recognition methods are analyzed, and a method of handwritten Chinese character recognition based on depth belief network fusion model is proposed.Firstly, the modified quadratic discriminant function classifier is used to recognize the simpler Chinese characters, and the more complex Chinese character images are processed by the depth belief network model.The specific process of division of labor coordinates the division of labor between the two classifiers in the recognition task by defining the credibility, so as to achieve the ideal recognition effect.Then, the convolution neural network is better than the traditional image classification method in feature extraction, so the convolution neural network is applied to the off-line handwritten Chinese character recognition problem.The advantages and disadvantages of convolution in handwritten Chinese character processing are analyzed. Aiming at the problem of similar handwritten Chinese character classification, the convolution neural network is further improved to make full use of the advantages of feature extraction in the network. The second classifier uses the classical SVM classifier.The combination of the two is better for the recognition and classification of similar handwritten Chinese characters.In this paper, MNIST handwritten, CASIA-HWDB1.1 collected by Institute of Automation, Chinese Academy of Sciences, and HCL2000 data set collected by Beijing University of posts and Telecommunications are selected.Experimental results show that the proposed method based on convolution neural network can achieve good recognition effect.
【学位授予单位】：哈尔滨理工大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.43;TP181

【相似文献】