摘要
提出一种基于B样条概率密度函数(PDF,probability density function)估计的复杂纹理图像分类识别方法,主要包括图像纹理PDF特征学习、表征以及纹理类型识别3个步骤。在图像纹理PDF特征学习和表征中,引入各向异性高斯导数方向滤波器获得纹理图像多尺度和多方向空间结构;然后基于预先固定的B样条基函数,将图像空间结构PDF估计转化为与基函数相对应的权值向量估计;之后采用最远邻聚类方法,获得图像空间纹理结构的PDF特征字典库;最后采用最近邻方法,获得各类纹理在特征字典库上的直方图分布表示。在纹理类型识别阶段,基于直方图距离测量结果实现纹理图像分类识别。在不同纹理图像数据库上进行了大量的验证性和对比性实验,实验结果表明所提方法的有效性和优越性。
A B-spline probability density function (PDF) estimation-based texture image identification method is presented in this work, which can be categorized into three processing phases,including PDF feature dictionary learn ing phase,PDF feature representation phase and texture identification phase.In the PDF feature dictionary learning and feature represe ntation phases,anisotropic Gaussian derivative filters are adopted to achieve the image′s spatial textural structures under various Gaussi an observation scales and orientations in advance.Next,the non-parametric PDF feature of the local spatial-structural texture is transfor med to parametric weighting vector estimation based on the pre-fixed B-spline basis functions.The PDF feature dictionary base of the tex ture image samples is obtained by the furthest-neighbor clustering method,and the nearest neighbor method is used successively to obtain the histogram distributions for all kinds of image texture categories for the following texture identification.Finally,the image texture is identified by the simple histogram distance measurement. Abundant confirmative and comparative tests on the commonly-used testing sets i ndicate the effectiveness and superiority of the proposed method .
出处
《光电子.激光》
EI
CAS
CSCD
北大核心
2017年第5期538-546,共9页
Journal of Optoelectronics·Laser
基金
国家自然科学基金(61501183
61472134)
湖南省科技计划基金(2013FJ4051)资助项目
关键词
B样条概率密度估计
纹理图像分类
各向异性高斯核
最远邻聚类
图像统计建模
B-spline probability density estimation
texture image classification
anisotrophic Gaussiankernel
furthest-neighbor clustering
image statistical modeling