基于Contourlet变换和局部二值模式图像纹理分类研究及其应用

发布时间：2018-09-12 09:50

【摘要】：随着计算机技术和人工智能的不断普及,图像的应用领域非常广泛,其中图像特征提取占据重要的主导地位,也是当前计算机视觉领域中热门的研究课题之一。在过去的研究中,局部特征的深入研究推动了整个计算机视觉领域的快速发展,其中最著名的局部二值模式(LBP)是局部特征中一种非常简单高效的局部描述子。自LBP提出来之后,受到众多研究者的追崇,改良的算法被应用于各种计算机视觉和模式识别领域中,包括纹理分类、人脸识别、目标检测等。但是各种LBP算法仍会存在一定的缺陷和瓶颈,如LBP构造的直方图向量长度过长,旋转不变能力不够突出,噪声鲁棒性还不够强等问题。为了增强LBP的鉴别能力,提升抗噪声的鲁棒性,本文对纹理分类中所用到的关键技术进行了深入研究后,提出了一些新算法,并应用到特定的纹理数据库进行验证。另外,Contourlet变换继承了Curvlet变换的各向异性的多尺度关系,能够提取图像的内在几何特征,因此本文也提出了基于Contourlet变换的纹理分类方法,并应用于纸币识别中。本文的主要工作和贡献如下:1、系统研究了纹理分类中较著名的纹理特征提取算法,如LBP、CLBP、DLBP、SCLBP等局部二值模式描述子,详细陈述了它们的工作实现原理,分析它们的优势和需要改进的地方,并介绍它们所应用的典型领域。2、研究了局部二值模式的改进算法。针对LBP的鉴别能力和抗噪能力较弱,本文在前人研究的基础上,对BRINT的提取识别算法进行改进,提出一种新的基于BRINT的尺度不变特征提取方法,通过BRINT的尺度不变空间分析,可以得到最优的尺度不变特征,并用不同的纹理数据库进行验证,得到较好的效果。首先,LBP描述子解决了旋转不变和灰度不变特性,但在尺度不变特性方面没有得到很好解决。我们提出的方法正好解决了这个瓶颈。其次,不同于其他传统的局部尺度不变特征,我们不需要估计局部的尺度,而仅仅使用了全局描述子去实现尺度不变特性。最后,我们将直方图横跨不同尺度空间,最终取每个分量最大值,并整合成一个最优直方图,实现尺度不变特性。这种新的方案在纹理数据库中能得到一定的提升效果。3、研究了基于Contourlet变换和统计特性的纸币纹理分类。首先介绍纸币识别的基本现状,详细介绍contourlet变换的工作原理以及相关的变种,随后利用contourlet变换分解出来的频率子带进行模式统计特征提取,结合灰度共生矩阵方法,采用基于支持向量机(SVM)方法实现了纸币纹理分类。最后与相关的方法比较,验证本文算法的可行性。
[Abstract]:With the popularization of computer technology and artificial intelligence, the application of image is very extensive. Image feature extraction plays an important role in the field of computer vision, and it is also one of the hot research topics in the field of computer vision. In the past, the in-depth study of local features has promoted the rapid development of the whole field of computer vision. The most famous local binary pattern (LBP) is a very simple and efficient local descriptor in local features. Since LBP was proposed by many researchers, the improved algorithm has been applied to various fields of computer vision and pattern recognition, including texture classification, face recognition, target detection and so on. However, there are still some defects and bottlenecks in various LBP algorithms, such as long histogram vector length constructed by LBP, insufficient rotation invariant ability, and insufficient noise robustness. In order to enhance the discriminative ability of LBP and enhance the robustness of anti-noise, the key techniques used in texture classification are deeply studied in this paper, and some new algorithms are proposed, which are applied to a specific texture database for verification. In addition, Contourlet transform inherits the anisotropic multi-scale relation of Curvlet transform and can extract the intrinsic geometric feature of image. Therefore, a texture classification method based on Contourlet transform is proposed and applied to paper currency recognition. The main work and contributions of this paper are as follows: 1. The famous texture feature extraction algorithms in texture classification, such as local binary pattern descriptors such as LBP,CLBP,DLBP,SCLBP, are systematically studied, and their implementation principles are described in detail. This paper analyzes their advantages and needs to be improved, and introduces the typical field. 2, and studies the improved algorithm of local binary pattern. Aiming at the weak discriminant ability and anti-noise ability of LBP, this paper improves the extraction and recognition algorithm of BRINT on the basis of previous research, and proposes a new scale-invariant feature extraction method based on BRINT, which is analyzed by scale invariant space of BRINT. The optimal scale invariant features can be obtained and verified by different texture databases, and good results can be obtained. First, the LBP descriptor solves the rotation invariance and gray invariance, but it is not well solved in the scale invariance. The method we put forward just solves this bottleneck. Secondly, unlike other traditional local scale invariants, we do not need to estimate local scales, but only use global descriptors to realize scale invariants. Finally, we take the maximum value of each component across different scale spaces, and integrate the histogram into an optimal histogram to realize the scale-invariant property. This new scheme can get a certain improvement effect in texture database. The paper studies the paper currency texture classification based on Contourlet transform and statistical characteristics. This paper first introduces the basic status of banknote recognition, introduces in detail the working principle of contourlet transform and its related variants, then extracts the statistical features of patterns by using the frequency subbands decomposed by contourlet transform, and combines with the method of gray level co-occurrence matrix. The paper currency texture classification is realized based on support vector machine (SVM) (SVM) method. Finally, compared with the related methods, the feasibility of the algorithm is verified.
【学位授予单位】：广东工业大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】