This procedure was not tested for algorithmic complexity, but was compared to a simple GLCM implementation and was found more stable and faster。

The implementation of the Advanced Texture Heuristic module was performed in C++, as the MSEG algorithm。 The modified version of the algorithm was called Texture-based MSEG。

3。RESULTS AND DISCUSSION

The implemented version of the MSEG algorithm was tested on a variety of image data, in order to assess the quality of the results, its ability to be generic and its speed。 Evaluating the results of a segmentation algorithm does not depend on the delivery of semantic objects, but rather on the generation of good object primitives useful to further classification steps。

The algorithm was designed to provide over-segmentation so that merging of segments, towards the final image semantics, to be achieved by a follow up classification procedure。 Boundary distinction and full-scene segmentation were of great significance。 Since the eCognition software (eCognition User Guide 2005) is greatly used for object oriented image analysis purposes the evaluation of results was mainly based on comparison with outputs from eCognition。 Also, a comparison was made to the Simple Profile results of MSEG, to show how the texture features performed along with region merging segmentation。

For the evaluation of the algorithms a Landsat TM image was used。 The eCognition software was used to provide segmentations with scale parameters 10 and 20。 The color criterion was used with a weight of 0。7 and the shape criterion with weight 0。3。 The results are shown in Figures  3  and 4。 Then, the simple profile of MSEG was used to provide segmentations with scale parameters of 400 and 700 respectively, to simulate the mean object size of eCognition’s results。 It should be noted that scale parameters are not compatible between the two algorithms, but are implementation dependent。 MSEG results are shown in Figures 5 and 6。

In Figures 7, 8 and 9, the results from the texture-based MSEG are shown。 Similar scale parameters with MSEG’s simple profile results have been used, and also the same weights for color and shape criteria。 In Figure 7, a scale parameter 400 and the texture parameter of 2。0 were used。 In Figure 8, scale parameter was the same, but texture parameter was 1。0。 Finally in Figure 9, scale parameter was set to 2500 and texture parameter was set to 3。0。

关键词：基于对象的图像分析，遥感，共生矩阵，角秒矩

摘要：本研究的目的是设计和开发基于区域的多尺度分割算法复杂纹理特征的整合，为面向对象的图像分析提供一种低层次的处理工具。这个实现的算法称为基于纹理的MSEG，可以描述为一个区域的合并过程。该算法由双型材组成。简单的来说，分割算法的主要部分已经包括在里面。第一个对象表示是图像的单像素论文网。通过迭代成对对象的融合，这是在多次迭代后，调用传递，实现最后的分割。对象合并的标准是一个同质化的成本度量，定义为对象异质性，并计算每个可能的基于光谱和形状特征的对象合并。异质性的尺度参数和用户定义的阈值相比，是用来确定合并的。订单的原始对象通过一个过程定义（起点估计），基于图像分割的统计指数和抖动算法来处理。高级配置文件被实现为一个扩展的简单的概要文件和包括全球为提高切割功能多分辨率功能异质性启发式模块。作为先进的配置文件的一部分，一个集成的纹理特征对该地区合并分割过程通过一个先进的纹理启发模块实现。对这个texture-enhanced分割方法，复杂纹理的统计措施必须计算基于对象，而不是正交图像区域。为每一个图像对象级灰度共生矩阵及其统计特性计算。先进的纹理启发式模块，集成新的启发式决策的对象合并，涉及相邻图像的相似性度量对象，基于计算纹理特性。在c++中实现的算法，并测试了不同传感器的遥感图像，决议和复杂水平。产生的原始对象与那些其他分割算法相比，结果是令人满意的。简单轮廓的原始对象和基于纹理基元的对象比较还显示，纹理特征可以提供良好的分割结果，除了光谱和形状等特点。