基于随机投影的图像纹理特征及其应用研究

As one of the fundamental characteristics of images, texture plays an important role in the areas of image analysis and pattern recognition. The question of how to extract simple, yet highly efficient texture features robust to imaging environment changes including both geometrical and photometric, such as changes in illumination, image rotation, view point and scale, is therefore still acute and of high importance in many image analysis application and has received considerable attention during the past decades. The purpose of this project is the development of texture feature, based on the random projection (RP) technique. The two application areas studied in this project are texture classification, material categorization and segmentation. The overall goal of this work is to explore simple, low dimensional, highly effective and robust texture features, which are nevertheless capable of encapsulating different types of complimentary texture information, to develop feature combination approaches to combine the proposed texture features, to design texture classification framework and segmentation framework with the proposed single or combined texture features and to increase the robustness and universality of their application in multiple applications. To this end, we intend to conduct research in the following three subjects: (1) In the research of texture feature extraction, by leveraging the sparse nature of texture images and inspired by theories of sparse representation and compressed sensing, we introduce random projection technique into the field of texture classification, and propose unconventional RP texture features, and sorted random projection (SRP) rotation invariant texture features. Along with this work, we intend to develop approaches to increase the robustness to noise of the proposed random features; (2) In the research of feature combination and classifier, motivated by the excellent texture classification performance of the proposed RP and SRP features, this project seeks to build an application-oriented adaptive feature combination system, which is capable of identifying relevant class specific aspects while being robust to intra-class variations, by combining the proposed random features based on support vector machines and multiple kernel learning, aiming at learning the correct weighting of different features from training data; (3) In the research of potential application, based on the previous research, this work develops simple, highly efficient, yet robust texture classification system, material recognition system and fast texture segmentation system based on the proposed single random feature or multiple random feature combination.

纹理是图像的主要特性之一，在图像分析与模式识别领域占有重要地位。如何简单、高效提取具有光照、旋转、视点和尺度不变性的纹理特征，实现高性能图像分析一直是人们关注的研究热点之一。本课题以纹理分类和分割为主要研究对象，研究基于随机投影的纹理特征理论和方法，旨在提取简单、低维、有效、稳健的多类型新纹理特征，建立组合这些新特征进行纹理图像分类与分割方法，提高处理性能及适应多应用领域的能力。本课题拟开展如下研究：（1）在特征提取方面，利用纹理图像的稀疏性，引入随机投影和压缩感知理论，提出新随机纹理特征和有序随机投影旋转不变纹理特征，并研究增强其抗噪性的方法；（2）在特征组合与分类器方面，在所提新纹理特征基础上，设计面向应用的基于支持向量机和多核学习的自适应最优特征学习方法；（3）在应用方面，构建基于随机投影或组合随机特征的高效、稳健的纹理分类和材质识别系统以及研究纹理分割框架的设计及其快速实现。

本项目从基础理论和实现方法两个角度对鲁棒纹理分类和纹理分割问题展开了研究，主要研究成果包括：.一、首次将随机投影技术和Johnson-Lindenstrauss引理引入纹理分析领域，设计了一种简单、新颖而又有效的RP分类器。新的RP分类器利用了纹理图像具有稀疏性这一潜在的先验知识，其特征提取过程简单直观，复杂度低，可扩展性好，稳健且十分便于工程实现。.二、针对旋转不变性纹理图像分类问题，在RP分类器的基础上，提出一种新的有序随机投影（ SRP）分类方法。提出了三个基于强度信息的随机特征描述子：SRP Global、SRP Square和SRP Circular，以及两种基于强度差分的新局部描述子：SRP径向差分描述子（SRP Radial-Diff）和SRP角向差分描述子（SRP Angular-Diff）。.三、为进一步提高纹理特征的稳健性以及对纹理图像成像环境的适应能力，深入分析和挖掘不同SRP随机纹理特征之间的内在关联，由此提出一种基于多核学习的多种SRP特征融合新纹理分类系统。将SRP Circular、SRP Radial-Diff和SRP Angular-Diff进行组合以获得更全面的纹理信息表示。提出通过多核SVMs对它们进行组合，进一步提高了纹理分类的正确率与稳健性。.四、针对纹理分割问题，以构建实用化的纹理分割框架为目的，研究了随机投影在纹理分割中的应用，提出了基于随机投影和单一BoW的纹理分割新方法。提出了一种基于局部径向差分算子、随机投影与BoW模型的简单、有效且稳健的有监督纹理分割新方法。.五、首次探索实用性强的LBP方法和理论性扎实的压缩感知方法之间的联系，提出几种新的理论简单而又有效的多分辨扩展LBP纹理特征描述子（NI-LBP, CI-LBP, RD-LBP和AD-LBP），并对它们进行联合，用于光照不变和旋转不变纹理图像分类，分类性能相比传统LBP方法有了大幅度提高。.六、针对含有噪声的纹理图像的分类问题，提出两种快速而又有效抗噪局部纹理特征描述子：BRINT和MRELBP。在提出的两种新的鲁棒描述子BRINT和MRELBP中，相比而言，MRELBP的性能更加优秀，就特征提取计算复杂度、特征维数、分类精准度和对图像旋转变化、尺度变化和光照变化以及噪声污染几个指标综合考虑，MRELBP为现有的最优秀的纹理特征描述子。