【作者】 张军；

【导师】 梁继民；

【作者基本信息】 西安电子科技大学 ， 模式识别与智能系统， 2014， 博士

【摘要】 纹理分析是计算机视觉领域一个古老却经久不衰的研究方向,而纹理表征也是诸多应用中的基本环节。纹理所包含的局部结构既拥有准周期重复又有随机出现的特性,因而统计的表征成为长久以来纹理分析的一个研究热点。如何精确并且有辨识力地描述局部结构和如何对局部结构描述子进行矢量量化是统计表征中最重要的两个问题。针对这两个基本问题,本文的主要研究内容包括以下几部分:(1)局部描述子的生成和矢量量化方法研究。针对如何生成局部描述子的问题,提出了基于局部方向能量分解的策略(局部能量模式,LEP),采用可分解二阶高斯导数滤波器快速滤波,并结合非线性变换和有偏归一化达到亮度和对比度的不变性。然后针对所提出的局部描述子,采用N进制编码的策略进行矢量量化。在进行N进制编码过程中,提出了一种采用直方图均衡化策略来训练自适应量化阈值的方法,从而比经验阈值更具有拓展性。此外,利用选择能量主方向和合并直方图条目的方式得到旋转不变性的表征,并且通过金字塔表征结合错位计算距离的方式达到尺度不变性的纹理分类。该方法在传统纹理数据库上具有良好的分类性能,并被初步应用于全天空极光图像数据的分类,为海量极光图像的自动分类奠定了基础。(2)局部描述子的生成。针对基于方向分解的表征方法大都只得到离散性旋转不变性的问题,提出了两种连续性旋转不变性描述子。第一种描述子采用一阶和二阶高斯滤波器最大响应描述局部结构(连续性最大响应描述子,CMR),第二种描述子采用曲面主曲率组合进行局部描述(主曲率描述子,PC)。然后采用基于纹元字典(码本)的矢量量化方法对纹理图像进行表征。理论分析和实验结果都表明这两种描述子具有良好的旋转不变性能,该算法不仅在未旋转的纹理分类中拥有稳定的性能,而且在旋转纹理的分类中取得了明显高于其他主流方法的结果。通过大量实验分析发现,连续性最大响应描述子具有更好的类间区分能力,而主曲率描述子拥有更好的类内凝聚能力。(3)局部描述子的生成。针对传统基于尺度错位和分辨率错位计算尺度不变距离效率较低和表征维度过高的问题,提出了一种尺度不变的纹理表征方法。该方法联合图像频率分解和梯度方向两种不同的尺度不变量来共同描述局部结构。通过采用多个方向的频域楔形滤波器响应进行频率分解,提出了一种新型的投票方式来获取图像在离散坐标系下的尺度不变梯度方向。该方法基于尺度不变量来描述局部结构,因而表征本身具有尺度不变性,不需要错位计算尺度不变性距离。该算法在尺度改变的纹理分类中得到了更加优异的结果。(4)矢量量化方法研究。针对传统矢量量化的表征维度过高和效率较低的问题,提出了一种快速的矢量量化方法。该方法通过结合纹元字典(码本)离线训练一个映射查找表,重新标记N进制编码生成的编码数字,达到以较低维度快速而准确的标记局部描述子的目的。相比于基于码本的矢量量化方法,所提出的方法提供了一种更快速地寻找最近邻纹元的策略,由于采用了查找表映射的方式,这种方法不再需要计算局部描述子与纹元之间的距离;相比于N进制编码的方法,所提出的方法采用对编码再标记的方式来降低表征维度,并且可以对更高维度的局部描述子进行矢量量化。结合多种典型的局部描述子进行纹理表征,在分类实验中拥有十分稳定的分类性能。(5)动态纹理表征。针对动态纹理图像序列的分析问题,将局部能量模式从二维静态表征扩展到三维动态纹理的表征,提出了体局部能量模式(VLEP)表征方法。该方法具有较高的执行效率,在UCLA动态纹理库的识别实验中得到了非常高的识别性能,识别率明显高于其他经典的动态纹理表征方法。更多还原

【Abstract】 Texture analysis is an old but long lasting research area for computer vision, and the texture representation is an important step for most applications. Texture description,unlike other forms of representation where the objects have a definite structure, most textures have large stochastic variations and quasi-periodicity structures which make them difficult to model. Therefore, local statistical representation is a hot topic for texture representation. In order to obtain the statistical representation, each local structure should be labeled distinctively. Specifically, two fundamental problems to label the local structures are: first, how to generate local descriptor to describe local structures with higher distinguishing ability, lower redundancy and imaging conditions invariance; second, how to quantize the local descriptor distinctively(by vector quantization) to preserve more structure information. Addressing to the two problems, our research focuses on the following parts:(1) Addressing to the problem of how to generate local descriptor, we use normalized local-oriented energies to generate local descriptor, which describe the local structures distinctively and are less sensitive to imaging conditions. Then, addressing to the problem of vector quantization, each local descriptor is quantized by self-adaptive quantization thresholds determined in the learning stage using histogram equalization, and the quantized local feature vector is transformed to a number by N-nary coding, which helps to preserve more structure information during vector quantization. Finally, the frequency histogram is used as the representation feature. Furthermore, according to the selection of dominant orientation and combination of histogram bins, the representation achieves rotation invariance. By using pyramid representation and shifting strategy for calculating dissimilarity, the proposed method obtains the scale invariance. The performance was benchmarked by material categorization on KTH-TIPS and KTH-TIPS2-a databases. The proposed method was compared with typical statistical approaches. The results show that our method is superior to other methods on the KTH-TIPS2-a database, and achieving competitive performance on the KTH-TIPS database. Besides, we preliminary attempt to use the proposed method for auroral image classification. The results lay the foundation for automatical classification of massive auroral images.(2) Texton dictionary-based texture representation approaches have been proven to be effective for texture classification. We propose two types of local descriptors based on Gaussian derivatives filters, both of which have the property of continuous rotation invariance. The first descriptor directly uses the maximum of the filter responses named as continuous maximum responses(CMR). The second descriptor rectifies the filter responses to calculate principal curvatures(PC) of the image surface. The texton dictionary is learned from the training images by clustering the local descriptors, and the representation of each image is the frequency histogram of the textons. The classification results compared with some other popular methods on the CUReT, KTH-TIPS and KTH-TIPS2-a datasets show that representation based on CMR achieves best classification result on the CUReT dataset. The representation based on PC achieves the best classification results on the KTH-TIPS and KTH-TIPS2-a datasets, and the classification performance is robust on different datasets. The experiments of rotation invariant analysis implemented on the Brodatz album illustrate that the CMR descriptor has good inter-class distinguish ability and PC descriptor has strong intra-class congregate ability. The results demonstrate that the proposed local descriptors achieve remarkable performance to classify the rotated textures.(3) This paper proposes an effective scale invariant texture representation based on frequency decomposition and gradient orientation. First, the image intensities are decomposed into different orientations by using the wedge filters in frequency domain, and the N-nary coding method is adopted for the vector quantization. Second, the scale invariant gradient orientation is generated by selecting the most stable value of the gradient orientation with different Gaussian scales. Finally, the 2D joint distribution of the two types of local descriptors is used as the representation. The performance was evaluated by the texture classification using nearest neighbor classifier. Simple but not ordinary, our method achieves the state of the art classification performance on the KTH-TIPS dataset under the traditional experimental design. Moreover, the main experiments were conducted on the KTH-TIPS and KTH-TIPS2-b datasets with the experimental designs of scale invariance validation. Compared with the methods of basic image features(BIFs) and local energy pattern(LEP), the proposed representation achieves superior performance with a much lower dimension of representation.(4) Addressing to the problem of vector quantization, we propose a fast vector quantization(VQ) method for statistical texture representation by combining two approaches of codebook and N-nary coding. Compared with the codebook, the proposed method offers a more efficient way of finding the nearest texton for the local descriptor, which does not need to calculate the distances between textons one by one, since a mapping procedure avoids the time-consuming calculation of the codebook. Compared with the N-nary coding, the proposed method relabels the coding numbers, which reduces the dimension of representation and makes it possible to quantize relatively higher dimensional local descriptor. In the experiment of texture classification, by using different sorts of local descriptors, the classification performance of proposed method does not decrease too much,on a contrary, it is more moderate and stable. Above all, it is estimable for our method to obtain stable performance with high efficiency and low dimension simultaneously, which suggests its practical application value.(5) Since the local energy pattern(LEP) achieves competitive classification performance for two dimensional texture image, we extend the method for three dimensional dynamic texture representation motivated by three dimensional local binary pattern. We extract the volume LEP(VLEP) by considering the dynamic texture sequence as three dimensional data.The 3D steerable filters replace 2D steerable filters to capture the texture and motion information simultaneously.更多还原