ABSTRACT:The objective of this research was the design and development of a region-based multi-scale segmentation algorithm with the integration of complex texture features, in order to provide a low level processing tool for object-oriented image analysis。 The implemented algorithm is called Texture-based MSEG and can be described as a region merging procedure。 The algorithm is composed of two profiles。 In the simple profile, the main part of the segmentation algorithm was included。 The first object representation is the single pixel of the image。 Through iterative pair-wise object fusions, which are made at several iterations, called passes, the final segmentation is achieved。 The criterion for object merging is a homogeneity cost measure, defined as object heterogeneity, and computed based on spectral and shape features for each possible object merge。 The heterogeneity is then compared to a user defined threshold, called scale parameter, in order for the decision of the merge to be determined。 The processing order of the primitive objects is defined through a procedure (Starting Point Estimation), which is based on image partitioning, statistical indices and dithering algorithms。 The advanced profile was implemented as an extension of the simple profile and was designed to include multi-resolution functionality and a global heterogeneity heuristic module for improving the segmentation capabilities。 As part of the advanced profile, an integration of texture features to the region merging segmentation procedure was implemented through an Advanced Texture Heuristics module。 Towards this texture-enhanced segmentation method, complex statistical measures of texture had to be computed based on objects, however, and not on orthogonal image regions。 For each image object the grey level co-occurrence matrices and their statistical features were computed。 The Advanced Texture Heuristics module, integrated new heuristics in the decision for object merging, involving similarity measures of adjacent image objects, based on the computed texture features。 The algorithm was implemented in C++ and was tested on remotely sensed images of different sensors, resolutions and complexity levels。 The results were satisfactory since the produced primitive objects, were comparable to those of other segmentation algorithms。 A comparison between the simple profile derived primitive objects and the texture based primitive objects also took place showing that texture features can provide good segmentation results in addition to spectral and shape features。80801

1。INTRODUCTION

1。1Recent developments in Remote Sensing

Remote sensing has recently achieved great progress both in sensors and image analysis algorithms。 Due to very high resolution imagery, such as IKONOS and Quick Bird, traditional classification methods, have become less effective given the magnitude of heterogeneity appearing in the spectral feature space of such imagery。 The spectral heterogeneity of imaging data has increased rapidly, and the traditional methods tend to produce classification errors such as multiple spectral signatures within a semantic object。 Those multiple signatures, cannot be effectively dealt with standard methods and tend to produce “salt and pepper” classification results, when one semantic object is composed of multiple spectral signatures。

Another disadvantage of traditional classification methods is that they do not use information related to shape, site  and spatial relation (context) of the objects of the scene。 Context information is a key element to photo-interpretation, and a key feature used by all photo-interpreters because it encapsulates expert knowledge about the image objects。 Such knowledge, however is not explicit, and needs to be extracted,    represented

and used for image analysis purposes。 In order to improve classification results from image analysis, it is of high importance to be able to use key features of photo- interpretation, such as shape and texture。