Keywords: Digital Photogrammetry, Feature Extraction, Knowledge Modelling, Image Understanding
Abstract
The study describes the procedure of digital image processing, and interpretation approaches to the sample image for object recognition. A classified section on an aerial image obtained by texture analysis is input as a potential residential area for further interpretation. Both linear (1D) and regional (2D) features are extracted from the classified section, based on image segmentation in terms of the degree to which their statistical, geometrical, and physical properties matched to rooftops and roads that mostly appeared in residential areas.
Linear features are detected first using Canny's method. The derived edge map is then defined for straight lines and vector information, such as line equation and intersection angle by the Hough Transform. Areal features are extracted and numbered through thresholding and coding techniques,based on relevantfactors, such as physical surfaces, size and scale, and housing density. These potential house blobs can also be located by computing the centroid thus providing vector information.
Rooftop extraction rate is studied by comparison of ground truth and roof blob number. An improved thresholding extracts 60% rooftops in the study. Knowledge required for the identification of rooftops and roads are discussed.
Representating and modelling for residential areas are based oncharacteristics of houses and streets, for example, roof chains and road intersections. Syntactic patterns that form the relations between identified features needs to be further studied. An early model of housing area is proposed and tested whether they fit the extracted features and relations between features for the purpose of image understanding.
Introduction
This research on digital photogrammetry studies some functions in computer vision technology and their performance when applied to photogrammetric mapping. It is based on Marr's vision theory of the primal sketch and image representation, with consideration to both physical (brightness) and psychological (human vision) factors in digital image analysis. The objectives are thus to develop the procedures and techniques that will support the processes of feature extraction and recognition of cartographic features in digital aerial images.
The process of segmentation partitions an input image into its constituent parts. The output of the segmentation stage usually consists of pixels constituting either the boundary of a region or all the points in the region itself. In either case, converting the raster data into a mathematical form suitable for computer processing is necessary. Three types of segmentation methods, histogram, discontinuity, and similarity, are used in practice.
Object representation is an essential process for transforming raw image data into a suitable geometric form for subsequent computer processing. Most cartographic features are usually identifiable in aerial photographs by a human interpreter. The geometric representation of objects in terms of their extracted linear and regional features can provide information for computer based description of the image contents.
The term "representation" in this study has two broad usages: one is to present different image processing stages for the mathematical description of objects, as indicated in [Marr 1982] and [Ballard and Brown 1982]; the other is to describe programmed human knowledge i.e. knowledge representation.
Marr [1982] stated that although human vision delivers a shape description of an object from one image, it is almost certainly impossible to do this in only one step, which leads to the idea of a sequence of representations, starting with descriptions that could be obtained straight from an image, but actually are carefully designed to gradually facilitate the subsequent recovery of an object's physical characteristics and mathematical properties about its shape, size etc.
Three vision representational stages are introduced by Marr [1982]:
1.
2D representation or
primal sketch --- the representation of the properties of a two-dimensional image, such as intensity changes and local two-dimensional geometry.
2.
2.5D representation --- the representation of the properties of visible surfaces in a viewer-centred coordinate system, such as the surface orientation, distance from the viewer, and discontinuities in these quantities; the surface reflectances; and some coarse descriptions of the prevailing illumination.
3.
3D representation --- an object-centred representation of the 3D structure and of the organisation of the viewed shape, together with some description of its surface properties.
The role of knowledge is displayed in Figure 1, showing four basic representational categories in computer vision.
Figure 1 The knowledge base of a complex computer vision system, showing the four basic representational categories. [Ballard and Brown 1982]
Pre-Processing
The significant topics included in pre-processing for this sudy are, the creation and comparison of multi-resolution images, and texture description and modelling, for residential areas. The image understanding system in the study provides a data structure for images used in different stages for specific projects. Texture measures derived by the co-occurrence matrix have resulted in a suitable method of classification on multi-resolution images [Chao and Trinder 1997]. Linear and areal features, such as rooftops and straight roads are extracted respectively for the purposes of segmentation, representation, and recognition.
Feature Extraction
'Features', as described in Sowmya and Trinder [2000], can have different meanings in photogrammetry and computer vision. In this paper , 'features' refer to lines, road crossings, corners, blobs, regions. The extraction procedure involves both linear (road boundaries) and regional (roof blobs) features. Roof groups form the main part of the process, but houses are not the only elements in residential areas. Important co-existing linear features, most probably straight roads, must also be considered.
