基于卷积神经网络的脱机手写汉字识别系统研究

发布时间：2018-02-14 14:45

本文关键词： 深度学习脱机手写汉字识别卷积神经网络 GPU Caffe　出处：《安徽理工大学》2017年硕士论文　论文类型：学位论文

【摘要】：脱机手写体汉字识别在模式识别的应用领域内是非常重要的分支,同时其也是计算机智能接口实现过程中的关键性技术。由于脱机手写体汉字类别庞大、字体结构复杂、相似汉字多、书写随意性大以及书写风格差异强等特点,其在模式识别领域中始终得到广泛的关注与研究,是众多学者研究的热点与难点。随着深度学习在机器学习领域中逐渐并快速发展,同时在模式识别部分方面也已取得了优秀的识别性能。深度学习中卷积神经网络的优势在于其可以自动获取学习样本的特征,能够避免手写汉字特征提取的难题,但其在训练样本时较为困难。本文结合卷积神经网络在脱机手写体汉字识别系统进行了相关的研究,针对训练样本困难问题进行了优化与改进,论文主要研究工作为:(1)通过研究和分析学习脱机手写汉字识别的相关技术与卷积神经网络的理论,验证了卷积神经网络在脱机手写汉字识别过程中的可行性;(2)通过HCL2000与CSAIA-HWDB两类型脱机手写体汉字数据库的比较,并对系统的硬件部分进行GPU架构设计,实现了GPU、CPU与SSD硬盘三者之间的数据信息交互的过程,结合Caffe软件平台的设计,实现了基于卷积神经网络的脱机手写汉字识别系统;(3)在实现的脱机手写汉字识别系统中,通过对卷积神经网络的结构优化,本文从卷积神经网络的激活函数、Dropout方法以及数据样本图像的弹性形变等方面提出四种改进的网络模型,并通过相应的实验研究表明,改进后的CNN_4卷积神经网络在脱机手写汉字识别过程中,其汉字识别准确率上提升了2.5%。由实验研究发现,本文改进后的卷积神经网络在脱机手写识别过程中,提升了网络模型的泛化能力、收敛速度以及识别准确率。
[Abstract]:Off-line handwritten Chinese character recognition is a very important branch in the application field of pattern recognition, and it is also a key technology in the realization of computer intelligent interface. It has been widely paid attention to and studied in the field of pattern recognition because of its many similar characters, great arbitrariness in writing and strong differences in writing style. It is a hot and difficult point for many scholars. With the gradual and rapid development of deep learning in the field of machine learning, The advantage of convolution neural network in deep learning is that it can automatically acquire the features of learning samples and avoid the problem of handwritten Chinese character feature extraction. However, it is difficult to train samples. This paper studies the offline handwritten Chinese character recognition system with convolutional neural network, and optimizes and improves the difficult problems of training samples. The main work of this paper is to study and analyze the related techniques of off-line handwritten Chinese character recognition and the theory of convolution neural network. The feasibility of convolutional neural network in off-line handwritten Chinese character recognition is verified. The comparison of two types of offline handwritten Chinese character databases, HCL2000 and CSAIA-HWDB, and the design of GPU architecture for the hardware part of the system are carried out. The process of data and information exchange between GPU CPU and SSD hard disk is realized. Combined with the design of Caffe software platform, the off-line handwritten Chinese character recognition system based on convolution neural network is realized in the off-line handwritten Chinese character recognition system. By optimizing the structure of the convolutional neural network, this paper proposes four improved network models from the aspects of the activation function of the convolution neural network, Dropout method and the elastic deformation of the data sample image. The improved CNN_4 convolution neural network improves the accuracy of Chinese character recognition by 2.5% in the offline handwritten Chinese character recognition process. From the experimental study, it is found that the improved convolution neural network is used in the offline handwritten recognition process. The generalization ability, convergence speed and recognition accuracy of the network model are improved.
【学位授予单位】：安徽理工大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.4;TP183

【相似文献】