光学学报
光學學報
광학학보
ACTA OPTICA SINICA
2009年
11期
3010-3017
,共8页
王雷光%刘国英%梅天灿%秦前清
王雷光%劉國英%梅天燦%秦前清
왕뢰광%류국영%매천찬%진전청
图像处理%纹理分割%均值漂移%Gabor滤波器%信息加权
圖像處理%紋理分割%均值漂移%Gabor濾波器%信息加權
도상처리%문리분할%균치표이%Gabor려파기%신식가권
image processing%texture segmentation%mean shift%Gabor filter%information weighting
高分辨率遥感影像呈现极其丰富的光谱和结构信息,传统的基于光谱的遥感影像分割方法往往使得分割区域过于细碎且分割精度不高.尝试将纹理信息引入到特征空间以期解决该问题.本文算法中,特征空间由光谱和纹理两类构成,并采用加权最小距离分类器.光谱信息通过对原始影像的变带宽均值漂移滤波获得,纹理信息由对原始影像逐波段采用多尺度伽博(Gabor)滤波器组滤波获得;依据训练样区中各特征维的方差确定该地物类别分类时特征维的权重,并通过训练样区的特征加权平均获得各地物类别的聚类中心;最后,将像素点归为到加权聚类中心距离最小的类别.实验结果表明,提出的均值漂移带宽确定方法是有效的,加权融合算法较基于光谱的分割方法在分割精度上有一定程度的提高.
高分辨率遙感影像呈現極其豐富的光譜和結構信息,傳統的基于光譜的遙感影像分割方法往往使得分割區域過于細碎且分割精度不高.嘗試將紋理信息引入到特徵空間以期解決該問題.本文算法中,特徵空間由光譜和紋理兩類構成,併採用加權最小距離分類器.光譜信息通過對原始影像的變帶寬均值漂移濾波穫得,紋理信息由對原始影像逐波段採用多呎度伽博(Gabor)濾波器組濾波穫得;依據訓練樣區中各特徵維的方差確定該地物類彆分類時特徵維的權重,併通過訓練樣區的特徵加權平均穫得各地物類彆的聚類中心;最後,將像素點歸為到加權聚類中心距離最小的類彆.實驗結果錶明,提齣的均值漂移帶寬確定方法是有效的,加權融閤算法較基于光譜的分割方法在分割精度上有一定程度的提高.
고분변솔요감영상정현겁기봉부적광보화결구신식,전통적기우광보적요감영상분할방법왕왕사득분할구역과우세쇄차분할정도불고.상시장문리신식인입도특정공간이기해결해문제.본문산법중,특정공간유광보화문리량류구성,병채용가권최소거리분류기.광보신식통과대원시영상적변대관균치표이려파획득,문리신식유대원시영상축파단채용다척도가박(Gabor)려파기조려파획득;의거훈련양구중각특정유적방차학정해지물유별분류시특정유적권중,병통과훈련양구적특정가권평균획득각지물유별적취류중심;최후,장상소점귀위도가권취류중심거리최소적유별.실험결과표명,제출적균치표이대관학정방법시유효적,가권융합산법교기우광보적분할방법재분할정도상유일정정도적제고.
High-spatial-resolution remote sensing imagery provides a large amount of spectral and structure information. However, their availability also poses challenges to conventional spectral segmentation methods, and the segmenation region is often too fragmentary and has low accuracy. In order to overcome this inadequacy, texture information is introduced into spectral feature space. In the algorithm, the new feature space consists of spectral and texture elements, and weighted minimum distance classifier is designed. Firstly, spectral feature is got by a variable bandwidth mean shift filtering procedure on original images, and texture feature is got by convolving original image with multiscale Gabor filter bank band by band. Secondly, the weight of certain feature dimension for a certain land class is determined by its deviation in the land class training area. Then, the clustering centre is also calculated by averaging weighted feature vectors in the training area. Finally, every pixel is classified into the class with nearest weighted distance. The experiments demonstrate that the presented band definition method using the variable mean shift filtering is effective and the combination of different features can achieve better performance than only using texture or spectral feature independently.