纹理图像特征提取与分类研究

发布时间：2018-06-25 02:31

本文选题：纹理 + 特征提取　；参考：《华东师范大学》2017年博士论文

【摘要】：纹理图像特征提取和分类在遥感、医学、农业、工业等领域有广泛应用,可以进行地形地貌检测、灾害预防、农作物监测、医学影像分析等。传统的纹理特征还存在一些不足,一些纹理特征对旋转、姿态、视角、尺度变化等较为敏感,分类时间较长,在一些实际应用中,这些纹理特征分类效果较差。针对纹理图像中旋转问题,本文提出一种新的多尺度旋转不变纹理特征(MSRIT)提取方法,MSRIT具有旋转不变性,可应用于旋转、视角、姿态、尺度变化等纹理分类。针对传统分类方法在纹理图像分类效率较低问题,本文提出了一种新的SVM分类模型(SVMpdip),并提出了针对该模型的求解方法:基于块消除法的原-对偶内点法(PDIPbe)。SVMpdip具有很高的分类准确率,且分类时间少于一些传统分类方法。MSRIT和SVMpdip方法可以处理实际应用中较为复杂的纹理分类问题。MSRIT方法是从多个尺度图像的多个旋转不变局部特征描述子中来提取图像纹理特征,这些纹理特征具有多尺度旋转不变性。在SVMpdip模型求解过程中,本文采用块消除法将中间过程系数矩阵分解为含有单位矩阵、对角矩阵等多个特殊矩阵的块矩阵,大大减少存储空间,减少了计算复杂度,提高分类效率。本文先寻找分类模型合适优化初始点,再求解分类模型,提高了模型求解收敛速度。在理论分析基础上,本文利用国际上典型纹理数据集进行了较多实验分析与评价,实验结果表明MSRIT分类准确度好于Gabor、GLCM、GLDM、LBP等传统纹理特征提取方法;SVMpdip 分类时间短于 SMO-P、SMO-K1、SMO-K2、CVX、quadprog、svmlight 等分类方法,效率高于它们,分类准确率也非常高。
[Abstract]:Texture image feature extraction and classification are widely used in remote sensing, medicine, agriculture, industry and other fields. They can be used for terrain and geomorphology detection, disaster prevention, crop monitoring, medical image analysis and so on. Some of the traditional texture features are sensitive to rotation, attitude, visual angle, scale change and so on, and the classification time is longer. In some practical applications, the classification effect of these texture features is poor. In this paper, a new multi-scale rotation invariant texture feature (MSRIT) extraction method is proposed to solve the rotation problem in texture images. MSRIT is rotation-invariant and can be applied to texture classification such as rotation, angle of view, attitude, scale change and so on. Aiming at the low efficiency of traditional classification methods in texture image classification, In this paper, a new SVM classification model (SVMpdip) is proposed, and a method for solving the model is proposed: the primal-dual interior point method (PDIPbe). SVMpdip based on block elimination method has high classification accuracy. The classification time is less than that of some traditional classification methods. MSRIT and SVMpdip can deal with the more complex texture classification problem in practical applications. MSRIT can extract image texture features from multiple rotation invariant local feature descriptors of multi-scale images. These texture features have multi-scale rotation invariance. In the process of solving SVMpdip model, the intermediate process coefficient matrix is decomposed into block matrices containing unit matrix, diagonal matrix and other special matrices by block elimination method, which greatly reduces the storage space and computational complexity. Improve the efficiency of classification. In this paper, the optimal initial point of the classification model is found first, and then the classification model is solved, which improves the convergence rate of the model. On the basis of theoretical analysis, more experiments are carried out with typical texture data sets in the world. The experimental results show that the accuracy of MSRIT classification is better than that of traditional texture feature extraction methods, such as Gabor-GLCM-GLDMU LBP, and the time of SVMpdip classification is shorter than that of SMO-PMO-K1 SMO-K2CVXMlight and other classification methods, such as SMO-PMO-K1CMO-K2CVXMlight, etc. The efficiency is higher than them, and the classification accuracy is very high.
【学位授予单位】：华东师范大学
【学位级别】：博士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】