非特定人脱机手写笔迹鉴别方法的研究

发布时间：2018-03-16 21:54

本文选题：笔迹鉴别　切入点：特征提取　出处：《华中科技大学》2016年硕士论文　论文类型：学位论文

【摘要】：笔迹鉴别是一种重要的人体生物特征识别方法,它在公安、司法、考古、金融和电子商务等各个领域都有广泛的应用,而非特定脱机手写笔迹鉴别是笔迹鉴别中应用范围最广的分支,是目前研究的热点和难点。本文主要研究非特定人脱机手写笔迹鉴别的算法。本文将在图像分类中常用的Bag of Words(BoW)方法运用到非特定人脱机手写笔迹鉴别中。在特征提取方面,我们对SIFT特征,NoSIFT特征,SURF特征,CNN激活特征,LBP特征进行详细介绍和讨论对比。受Contour-Hinge等特征的启发,文中提出基于轮廓点的ELBP和基于轮廓点的ESIFT特征,实验证明两种基于轮廓点的特征包含互补信息,将两种特征融合后可以进一步提高鉴别准确率。在特征编码层面,本文对传统的硬投票(Hard Voting)方法和LLC稀疏编码方法进行对比分析,首次提出将一种基于局部仿射子空间编码的方法(LASC)运用到笔迹鉴别。这种方法考虑到每个单词周围邻域空间信息,因此明显优于传统硬投票和LLC编码方法,当字典空间较大时,该方法不会过早的出现过拟合现象,随着字典空间变大,鉴别准确率可以进一步提高。同时,本文深入讨论分析了基于GMM的FV、UBM、KLD三种编码方法并进行了对比实验分析。之后本文对比分析了基于BoW的特征表达和基于GMM的特征表达各自优缺点以及各自性能。最后,本文提出一种多字典特征融合的非特定人脱机手写笔迹鉴别方法。通过将判别性较高的特征进行加权融合,在公开数据集ICDAR2013和CVL数据集上取得较好的效果。
[Abstract]:Handwriting identification is an important method of human biometric identification. It is widely used in many fields, such as public security, judicial, archaeological, financial and electronic commerce. Non-specific off-line handwriting identification is the most widely used branch of handwriting identification. This paper mainly studies the algorithms of off-line handwriting identification for independent people. In this paper, the Bag of Wordsof Bow method, which is commonly used in image classification, is applied to the off-line handwriting identification of independent people. In this paper, we introduce and compare the features of SIFT feature, surf feature and SIFT active feature. Inspired by Contour-Hinge and other features, we propose ELBP based on contour point and ESIFT feature based on contour point. Experiments show that the two features based on contour points contain complementary information, and the accuracy of the two features can be further improved after the fusion of the two features. In this paper, the traditional hard voting method and LLC sparse coding method are compared and analyzed. A local affine subspace coding method based on local affine subspace coding is proposed for the first time. This method takes into account the neighborhood space information of each word, so it is superior to the traditional hard voting and LLC coding methods, when the dictionary space is large. This method does not appear the phenomenon of over-fitting prematurely. With the dictionary space increasing, the discriminant accuracy can be further improved. At the same time, In this paper, three coding methods based on GMM are discussed and compared. Then, the advantages and disadvantages of feature expression based on BoW and feature expression based on GMM are compared and analyzed. In this paper, a new method of off-line handwritten handwriting identification for independent individuals with multi-dictionary feature fusion is proposed. By weighted fusion of the higher discriminant features, better results are obtained on the open dataset ICDAR2013 and CVL datasets.
【学位授予单位】：华中科技大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.41

【参考文献】