毕业论文

打赏
当前位置: 毕业论文 > 外文文献翻译 >

图像分割算法英文文献和中文翻译(3)

时间:2022-05-19 22:47来源:毕业论文
MSEG is basically defined as a region merging technique。 Like all algorithms of this kind, it was based on several local or global criteria and heuristics, in order to merge objects in an iterative

MSEG is basically defined as a region merging technique。 Like all algorithms of this kind, it was based on several  local or global criteria and heuristics, in order to merge objects in an iterative procedure, until no other merges can occur (Sonka et al 1998)。 In most cases, a feature of some kind (mean spectral values, texture, entropy, mean square errors, shape indices etc。) or combination of such features computes the overall “energy” of each object。 Various definitions of homogeneity (energy minimization  measures  or  measures  of  similarity  within   an

object) have been defined (Sonka et al 1998, Pal and Pal 1993)。 Recently, a very successful segmentation algorithm, embedded in the Object Oriented Image Analysis Software eCognition (Baatz & Schäpe 2000), implemented such measures of homogeneity, for making the merging decision between neighbouring objects, with very good results。 Some spectral and spatial heuristics were also used to further optimize the segmentation。 In the proposed segmentation algorithm, similar homogeneity measures were used, and then complex texture features were implemented in later stages。

Figure 1: Flowchart of the MSEG algorithm

In order for the MSEG algorithm to provide primitive objects, several steps of region merging (passes) were followed。 The purpose of the first segmentation pass (Figure 1) was  to initialize image objects and to provide the first over- segmentation, in order for the algorithm to be able to begin region merging at following stages。 Initially, the objects of the image are the single pixels。 During first pass, the algorithm is merging single pixels-objects pair wise, inside  each macroblock。 For the second pass of the algorithm (Figure 1), the objects created by the first pass were used in a new pair wise merging procedure。 Again, the same strategy of merging was used, finding the best match for each object, and then checking if there is a mutual best match in order to merge the two objects (Tzotsos and Argialas 2006)。 The Nth pass module, is called iteratively until the algorithm converges。  The algorithm is considered finished, when during the nth pass no more merges occur and the algorithm converges (Figure 1)。 Then, the objects are exported and marked as final primitives。

2。2MSEG algorithm – Advanced Profile Overview

The simple profile of the MSEG algorithm included the pass modules, as basic elements of a region merging segmentation procedure。  The  extension  of  the  Simple  Profile  was  used to

include extra functionality algorithms  and  innovative techniques for improving the results。 The Advanced Profile, as implemented at present, included the following modules:

the Multi-scale Algorithm (MA), and

the Global Heterogeneity Heuristics (GHH)

the Advanced Texture Heuristics

The Multi-scale Algorithm module was designed to give to the MSEG algorithm the ability to create multiple instances of segmentations for an image, with different scale parameters。 Thus, the produced primitive objects could vary in size and therefore, the multi-scale representation could model large image entities, as well as small ones。 In order to  include multiple instances of segmentation, inside an object-oriented image analysis system, those instances must be properly constrained to be integrated and used together (Tzotsos and Argialas 2006)。

The problem when dealing with multiple segmentations, is the compatibility between scales, in order to combine information and objects。 One simple way to deal with this problem is to create a multi-level representation, and incorporate the multiple segmentations within this representation, hierarchically。

But a single-level hierarchy is sometimes not flexible, when dealing with remote sensing classification problems (Argialas and Tzotsos 2004)。 A multi-level-hierarchy approach or a branch-based hierarchy model can represent more complex spatial relations。 Thus, in the present Multi-scale segmentation Algorithm, every new level depends only from the nearest (scale-wise) super-level or the nearest sub-level, or both (Tzotsos and Argialas 2006)。 图像分割算法英文文献和中文翻译(3):http://www.youerw.com/fanyi/lunwen_94045.html

------分隔线----------------------------
推荐内容