【作者】 夏瑜；

【导师】 岳丽华；

【作者基本信息】 中国科学技术大学 ， 计算机应用技术， 2014， 博士

【摘要】 随着互联网技术日新月异的快速发展,互联网数据大幅增长,人们生活的进入“大数据”时代。作为信息表达的一种方式,图像数据不需要使用过多的文字去描述,具有直观、信息量大等特点。伴随着图像信息规模的不断增长,图像处理技术被广泛应用于医疗、军事、农业、工业、服务业等各行各业中,如医学图像分析、军事目标检测与识别、植物形态特征测量、道路交通监控、三维立体建模、饮食餐厅服务推荐等。各个应用领域对于数字图像的需求不断增长,对于图像分析技术的要求也越来越高,如何有效分析和应用这些图像是一个非常重要的课题。作为描述图像的最有效手段,特征在图像处理中非常重要,目前常用的图像特征包括：颜色、形状、纹理等特征。其中,纹理特征作为人类视觉系统的一种对物体表面表现的感知形式,通过计算图像中像素点的灰度或颜色变化以及变化分布的规律特征,来反应物体表面粗糙度、方向性和物体表面符合的某种规则性。纹理分析技术被广泛应用于人脸识别、农产品质量监督、道路交通监控、遥感图像处理、基于内容的图像检索、机器人视觉等诸多领域,对于不同应用领域和不同图像类型,给图像纹理特征提取提出的需求不同,同时,由于纹理结构本身的复杂性和广泛性,使得纹理分析技术成为数字图像处理中的一个具有较高难度的科学领域。近些年来,众多研究者提出了不同的纹理特征提取方法,主要可以分为如下几个方面：基于统计的纹理特征提取、基于模型的纹理特征提取、基于结构的纹理特征提取和基于信号处理的纹理特征提取。由于具有计算量小、特征维度低、支持旋转不变性等优点,基于结构的纹理特征提取方法已经被广泛应用于各个研究领域和实际应用中,众多研究者取得了不错且较成熟的研究成果,但是传统基于结构的纹理特征没有考虑纹理的方向变化及空间分布特性,无法充分表达图像的纹理变化方向特征和纹理空间分布特征,在一些图像处理应用如图像检索中无法全面地描述不同类别的图像,也无法取得较好的检索结果。因此,深入研究结构纹理特征的提取方法,改进现有方法的不足具有很高的理论研究价值和应用前景,通过研究基于结构的纹理特征提取,可以提高纹理特征提取的有效性,也为各行各业的广泛应用提供理论基础。本文首先介绍了结构纹理特征提取研究的背景和意义,讨论了该领域的国内外相关工作,从理论上阐述了结构纹理特征提取的基本思想和研究思路。接着,论文围绕特征提取方法,分析传统结构纹理特征在纹理方向变化及空间分布特征提取上的不足,提出新的基于方向特征和空间分布的结构纹理特征描述子。在此基础上,提出基于该特征描述子的图像检索算法。最后,分析传统光学遥感图像舰船检测中的问题与不足,结合结构纹理特征改进遥感图像海陆分割结果,进而改进舰船检测的效果。本论文的主要贡献可归纳为以下几个方面：(1)针对局部二进制模式无法提取纹理方向特征的问题,提出了一种反映纹理方向特征的纹理特征描述子,通过计算像素在不同方向上的灰度变化模式,构造局部灰度变化的共生矩阵,最后通过统计不同灰度模式的变化均值和方差特征,补充局部二进制模式纹理特征中的方向特征和幅度特征,有效地提高了纹理特征描述的全面性；(2)针对传统的结构纹理特征描述子无法提取纹理空间分布特征,提出局部空间二进制模式和局部空间分布模式,并在此基础上提出多尺度局部空间二进制模式和完整的局部空间分布模式。局部空间二进制模式通过提取像素与像素间的灰度变化模式对,反映纹理变化在空间上的分布特征,并在多尺度条件下分析特征的全面性。局部空间分布模式计算像素与像素之间在不同方向上的灰度变化模式,完整的局部空间分布模式不仅在原图像上提取局部空间分布模式,同时在梯度图和滤波图上提取局部空间分布模式,来充分提取纹理特征,提高了纹理特征描述的完整性；(3)针对结构纹理特征具有的计算简单、运算量小、特征描述全面等优点,基于提出的结构纹理特征描述子,设计图像检索算法来验证结构纹理特征描述子的有效性与合理性,分别是基于多尺度局部空间二进制模式的图像检索算法和基于完整局部空间分布模式的图像检索算法。实验结果表明本文提出的算法较同类方法具有更优的检索结果,显著提高了图像检索的全局平均查准率、全局平均查全率；(4)针对传统的海陆分割算法分割效果的不足,提出一种基于局部二进制模式特征的海陆分割算法,通过计算灰度和局部二进制模式特征的综合特征图并分割,与传统的灰度图海陆分割结果结合,得到更优的海陆分割结果。实验结果表明该算法在保证舰船检测高正确率的情况下,大大降低了舰船检测的虚警率。本论文分析了目前结构纹理特征的不足,提出了基于方向变化和空间分布的结构纹理特征,加深了对结构纹理特征提取的研究,为结构纹理特征在不同领域中的应用提供了理论基础；同时分析了结构纹理特征在图像检索和光学遥感图像舰船检测中的应用,为结构纹理特征在更多领域中的进一步应用与发展拓展了思路。更多还原

【Abstract】 The wide application and rapid development of the Internet leads to the explosive growth of network data. As a commonly used way of information expression, images need not much words to describe the information contained in images. Images can be simply understood by human and contain lots of information. With the growing amount of image data, the image processing technology is widly used in medical, military, agriculture, industry, service, and other industries. Applications such as medical image analysis, military targets detection and recognition, plant morphological characteristics measurement, traffic monitoring,3D-modeling, restaurant recommendation, etc. The demand of various applications for digital images is becoming more exuberant and the requirements of image analysis technology are also getting more higher. How to analyze these images and apply them in different applications is an important topic.As the most effective way to describe the image, features are very important in image processing. Features such as color, shape and texture are the mostly used feature in digital image processing. Among these features, texture as a form of human visual system on the surface performance of the perception. Through the calculation of each pixel in the image grayscale or color variation and distribution characteristics of changes, to response the roughness of surface, in line with the direction characteristics and rules of object surface. Texture analysis technique has been widely used in face recognition, the agricultural product quality supervision, road traffic monitoring, remote sensing image processing, content-based image retrieval, robot vision and many other fields. For different applications and different image types, the requirements for image texture are different. At the same time, because of the complexity of texture structure itself, texture analysis technology has become a high difficulty of Science in the field of digital image processing.In recent years, many researchers have proposed various texture feature extraction methods, they can be divided into four main categories:statistical based methods, model based methods, structure based methods and the methods based on signal processing.Structure based methods has the advantages of small calculation amount, low dimensionality and rotation invariance, etc., has been widely used in various fields of researches and applications. Traditional texture features based on structure hasn’t consider the variation of texture direction and spatial distribution characteristics, which can not fully express the texture directional features and spatial distribution of texture features in image.In some applications such as image retrieval, these features not only can not correctly distinguish the different classes of images, but also unable to obtain better retrieval results. Therefore, the further study of structure based texture features extraction is very important, the improvement of the existing methods has very high research value in theory and application, which not only can improve the validity of the texture feature extraction, but also provide theoretical basis for different applications.This paper firstly describes the background and significance of structural texture features extraction, discusses the related field work and introduce the basic idea and research methods theoretically. Secondly, this paper focus on the method of feature extraction, analyze the shortcomings of traditional structural texture extraction methods on the texture variation direction and spatial distribution, propose the new structure texture feature descriptor based on directional characteristics and spatial distribution. Moreover, the image retrieval algorithms based on the feature descriptor are proposed. Finally, the analysis of traditional optical remote sensing image ship detection problems and shortcomings in the remote sensing image has been made, the improvement of sea land segmentation results combined with texture and structure is proposed in this paper to improve the effect of ship detection.The main contribution of this paper is shown as:(1) To solve the problem of unable to extract directional texture feature of local binary pattern, a new texture feature descriptor based on the texture directional variation is proposed. Through the calculation of pixel gray variation on different directions, the feature co-occurrence matrix is constructed to reflect the local gray-scale variation, finally through the statistics of different gray patterns, directional feature and amplitude characteristics are proposed to supplement the local binary pattern texture features, which can effectively improve the overall texture feature description;(2) Aiming at the problem of unable to extract the spatial distribution characteristics of texture, new structural texture feature descriptors named as local spatial binary pattern and local spatial distribution pattern are proposed. Based on these two descriptors, multi-scale local spatial binary pattern and completed local spatial distribution pattern are also proposed. Local spatial binary pattern based on the gray-scale variation pattern between pixels, and calculated to reflect the distribution of gray-scale variation. Moreover, with multi-scale considered, the texture feature can be expressed more comprehensively. Local spatial distribution pattern calculates the gray- level variation pattern between different pixels on different directions. Completed local spatial distribution pattern not only extract the local spatial distribution pattern of original gray image, but also extract the local spatial distribution pattern of gradient image and filtered image from the original iamge, which can improve the completity of texture feature extraction;(3) Based on the advantages of simple calculation, low computation and comprehensive feature description, applying the structral texture feature descriptors in image retrieval algorithms. To verify the effectiveness and validity, two image retrieval algorithms namely based on multi-scale spatial local binary pattern and completed local spatial distribution pattern are proposed. Experimental results show that, compared with the similar methods, the proposed algorithms has better retrieval results, which can both improve the average recall and precision results;(4) To improve the sea-land segmentation of traditional methods, this paper propose a new segmentation algorithm based on local binary patterns. By calculating the local binary patterns and the integrated feature map to segment the optical remote sensing images, and combined with the traditional gray sea land segmentation results, better segmentation results can be obtained. The experimental results show that the algorithm can ensure the high accuracy of the ship detection, which can also greatly reduce the false alarm rate of ship detection.In this research, the deficiencies of the current structural texture features are analyzed. Structral texture features of directional variation and spatial distribution are proposed to impress the research of structral texture feature extraction, which provides the theoretical basis for the texture and structure of applications in various fields. Moreover, the application of structural texture features in image retrieval and the ship detection of optical remote sensing image are discussed, which also provides a new way for more fields in the further application and development of structural texture features.更多还原