1。2Texture Image Segmentation and Object-based Image Analysis

Approaches have been developed the fields of Computer Vision and Remote Sensing, for texture analysis and image segmentation。 In addition to simple texture features, such as standard deviation and variance, Haralick proposed more complex texture features computed from  co-occurrence matrices (Haralick et al 1973, Haralick 1979)。 These second order texture features were used in image classification of remote sensing imagery with good results (Materka and Strzelecki 1998)。 Even more complex texture models have been used for texture modelling, classification and  segmentation, such as Hidden Markov Models, Wavelets and Gabor filters (Materka and Strzelecki 1998) with very good results in remote sensing and medical applications。 Several methods have been proposed for texture image segmentation, taking advantage of the  latest  texture  modelling  methods (Chen  et al 2002,  Fauzi

* Heroon Polytechneiou 9, 15780, Zografou, Athens, Greece。 Tel。 +302107722684

and Lewis 2003, Havlicek and Tay 2001, Liapis et al 1998)。 At the same time, image classification also moved towards computational and artificial intelligence methods (Sukissian  et al 1994, Benz et al 2004)。

During the last few years, a new approach, called Object- Oriented Image Analysis, integrated low level image analysis methods, such as segmentation procedures and algorithms (Baatz & Schäpe 2000), with high level methods, such as Artificial Intelligence (knowledge-based expert systems and fuzzy systems) and Pattern Recognition methods。 Within this approach, the low level image analysis produces primitive image objects, while the high level processing classifies these primitives into meaningful domain objects (Benz et al 2004)。

1。3Research Objectives

The main objective of this research was the integration of texture features into an object-oriented image segmentation algorithm。 It was desired that the modified segmentation algorithm could be used as a low level processing part of an object-oriented image analysis system so that to be applied at multiple image resolutions and to produce objects of multiple scales (sizes), according to user-customizable parameters。

A further objective was the ability of the produced algorithm to be generic and produce good and classification-ready results from as many remote sensing data as possible。 Remote sensing data are, in general, difficult to process, with complex textural and spectral information。 Therefore, there was a need for the algorithm to be able to handle texture information and context features in order to produce better segmentation results。

2。METHODOLOGY

2。1MSEG algorithm – Simple Profile Overview

The MSEG algorithm (Tzotsos and Argialas 2006)  was designed to be a region merging technique, since region merging techniques are fast, generic and can be fully automated (without the need of seed points) (Sonka et al 1998, Pal and Pal 1993)。 Given that existing Object-Oriented Image Analysis systems (eCognition User Guide 2005) have used such methods was also a strong argument for the effectiveness of the region merging techniques。

After the data input stage (Figure 1), an image partitioning procedure (Macroblock Estimation) was applied to the dataset resulting into rectangular regions of variable dimensions, called macroblocks。 Image partitioning was applied for computing local statistics and for computation of starting points。 Starting points were then used for initialization of the algorithm and for reproducibility of segmentation results。

After the Macroblock Estimation, the SPE module (Starting Point Estimation) computed local statistics and  provided starting points for initialization of the region merging process。 It should be stretched that starting points were not used as seed points (as in region growing techniques) but are used to keep track of the order in which all pixels were initially processed (Tzotsos and Argialas 2006)。