东篱科研大数据发现系统（DRDS）

位置：成果数据库 > 期刊 > 期刊详情页

基于多分辨共生矩阵的纹理图像分类

ISSN号：1000-1239
期刊名称：计算机研究与发展
时间：0
页码：1991-1999
分类：TP391.4[自动化与计算机技术—计算机应用技术;自动化与计算机技术—计算机科学与技术] TP391.41[自动化与计算机技术—计算机应用技术;自动化与计算机技术—计算机科学与技术]
作者机构：[1]智能感知与图像理解教育部重点实验室（西安电子科技大学）,西安710071
相关基金：国家自然科学基金项目（60970066,60972148,61077009,61072106,61001206）;北京市自然科学基金项目（7092020）;国家“八六三”高技术研究发展计划基金项目（2008AA012125）;教育部博士点基金项目（200807010003）;高等学校学科创新引智计划（111计划）基金项目（B07048）;中央高校基本科研业务费专项基金项目（JY10000902032,JY10000902001）
相关项目：基于Treelet变换的多时相SAR图像变化检测

关键词：纹理分类, 多分辨, 共生矩阵, 非下采样小波变换, 特征选择, texture classificationl multiresolution, co-occurrence matrix undecimated wavelettransformation feature selection

中文摘要：

共生矩阵是描述纹理特征的一种常用方法．首先提出一种新的特征提取算法——多分辨共生矩阵．多分辨共生矩阵是通过同时在非下采样小波变换的逼近子带和细节子带上提取共生矩阵来实现的，能够有机整合传统小波的多分辨特性和频谱信息，以及空域灰度共生矩阵的纹理结构信息．其次，分析了多分辨共生矩阵、灰度共生矩阵以及小波能量特征的物理意义，并从相关性出发提出了新的特征选择方法，有效地降低了特征维数．对标准纹理库的分类实验结果表明：多分辨共生特征对纹理具有更好的描述能力，其分类正确率超过小波能量特征、空域灰度共生特征：二者融合以及灰度梯度共生特征的结果；所提出的特征选择方法在降低特征维数的同时，能够保持分类正确率．

英文摘要：

Gray level co-occurrence matrix （GLCM） is widely used for the description of different textures because of its advantage of representing the texture structure. In order to hold the multiresolution property and spectrum information simultaneously, GLCM is often computed from each wavelet subhand, which, however, leads to much higher dimension of feature. To overcome this problem, a new feature extracting algorithm is proposed named multiresolution co-occurrence matrix （MCM）, whose feature vector has a comparative low dimension. For the task of dimension reduction, though the MCM is computed from each subband of undeeimated wavelet transformation as GLCM does, the parameters for the MCM are well chosen taking into consideration the scale and orientation of wavelet subbands. In addition, the feature dimension of the MCM can be further reduced through a new feature selection method based on the correlation analysis of inter-subbands and intra-subbands. Performance analysis is made in detail for the statistics of the MCM and wavelet energy feature. The proposed MCM is measured through the classification test on the Brodatz album. Experimental results demonstrate that MCM statistics outperforms other methods such as wavelet energy, GLCM, and the catenation of these two features. Extensive experiments on feature selection show the effectiveness of the proposed method, which can reduce the dimension successfully without any loss of classification performance.

同期刊论文项目