基于Gabor滤波方法的居民地识别分析

图像纹理对于高分辨率遥感图像的信息提取与目标识别具有重要意义。针对"北京一号"小卫星全色遥感图像非城市区域居民地块往往呈现出比较明显的方向性纹理的特点,扩充改进Gabor滤波方法进行提取。方法主要利用Gabor滤波器的多尺度、多方向滤波的性质,提取多尺度纹理特征集,并进行特征;而后利用多特征聚类实现图像的初步分割。由于分割是对特征进行聚类完成的,其结果可能存在一个居民地块由若干个相互间存在间隔的子区域组成、存在无用小斑块、居民地内部存在大量小孔洞等缺陷。针对上述不足,利用形态学尺度空间融合方法,对居民地块通过结构元素不断增大的闭运算进行迭代融合,并选择一个具备"最长生存期限"的类别个数作为最佳类数,选择首次出现该类别数的分割结果作为最后的识别结果。对延庆地区的小卫星影像进行了居民地提取,并与共生矩阵纹理分析方法进行了实验对比。结果表明方法是有效的,并在提取精度上具有优势。

Residential Area Recognizing with Gabor Filtering from High Spatial Resolution Remotely Sensed Imagery

Texture is one kind of important feature for information extraction or target recognition from high spatial remotely sensed imagery.In this paper,a Gabor filtering based method to recognize residential areas from remotely sensed imagery is proposed based on that it often takes on obviously directional textures of the rural residential areas on the panchromatic imagery of the Beijing-1 micro-satellite system.Part of the Gabor filtering method includes the following steps: 1)to extract multi-scale,multi-oriented texture features with the Gabor filter group,2) to rectify and smooth the features with feature filtering;and 3)to segment the image with k-means clustering.With these steps,the initial residential areas can be extracted but with many deficiencies which include the existence of interspaces,holes and useless patches within many residential areas mainly because they are only obtained with image clustering.To resolve these problems,a morphological scale space based method is used to dissolve these residential patches with iteratively enlarged closing operators.Firstly,a pair of closing,opening operators are used to remove all the noises.A scale space is then constructed by using closing operator with structuring elements of increasing size.By doing so,the connected components(the set of pixels with neighborhood relationships) in the image will merge into each other gradually and become a single cluster in the end.This is essentially a binary image segmentation process,and can also be treated as a hierarchical clustering. The final number of clusters is the one which survives(relatively,not absolutely) the longest scale range,and the clustering which first realizes this number of clusters is the most suitable segmentation.Several experiments are carried out to validate the method,which includes the comparison to the classical texture analyzing method of co-occurrence matrix.