el c í rculo de adentro
Capitulo 6 No hay salvaci ó n
This sub-section describes the image segmentation processes applied to the input satellite data. Due to the complexity of the segmentation process, as described above the process is illustrated in Figure 3.12. An example of a household satellite image obtained from Google Earth of the Site A is shown in Figure 3.12 (a). However, before household segmentation can be applied it was first necessary to register and align the image so that each household was aligned in a north-south direction. The purpose of this registration and alignment was to facilitate later
feature segmentation. The result is a shown in Figure 3.12(b), where the image shown in Figure 3.12(a) has been appropriately aligned.
The next step was to convert each RGB colour represented images into an sixteen colour indexed image (as shown in Figure 3.12(c)), and then to transform this into a greyscale im- age to which histogram equalisation can be applied. Recall that histogram equalisation is a method for image enhancement directed at ensuring an equal colour distribution [55, 63], (Fig- ure 3.12(d)), detail concerning histogram equalisation was presented in Chapter 2, Sub-section 2.3.1. The process commences by selecting a reference image which is then used for normali- sation purposes with respect to the remaining images.
Once the image enhancement process was complete the next stage was to segment the images so as to isolate individual households. It has already been noted that the households of interest are typically defined by a rough rectangular boundary in which buildings and related objects are located. The aim was to segment these images so that these rectangular areas can be clearly isolated, however the boundaries are frequently not well defined in that the edges are not continuous. Thus, for example, region-growing segmentation techniques would be unlikely to perform well, instead line (edge) segmentation was adopted as a more suitable form of segmentation for the given application domain.
More specifically, for the purpose of image segmentation the Canny edge detection algo- rithm [20] and the Hough Transform [42] were applied (see details above). Prior to applying Canny edge detection and the Hough transform, contrast adjustment was applied to the images so that the household boundaries could be more readily distinguished.
Canny edge detection was then applied, the result as shown in Figure 3.12(e); in the figure the detected edges have been highlighted. As a result of applying the Hough transform, we have a collection of “lines” as shown in Figure 3.12(f). Each line is defined by a start and end point, and aρandθvalue (length and direction).
The next part of the process is to fit a rectangle to this set of lines. This was achieved by applying a Least squares approach [18], applied to each group of lines approximating to the top, bottom, left and right sides of a rectangle (see the detail above). As a results the rectangle surrounding each household was demarcated by a pair of horizontal and a pair of vertical lines (Figure 3.12(g)). The intersections of the lines can then be found so as to delimit the surrounding rectangle in term of its four corners (Figure 3.12(h)). The boundaries found were then applied to the original image to determine the area of each individual household. The final result is a set of segmented household images such as that as shown in Figure 3.12(i). The segmentation process as applied to a Site B household image is shown in Figure 3.13. Figure 3.13(a) is the original image. Figure 3.13(b) is the image after Figure 3.13(a) has been appropriately aligned. Figure 3.13(c) is the sixteen colour indexed image. Greyscale trans- formation and histogram equalisation was then applied to give Figure 3.13(d). The resulting image after Canny edge detection has been applied is presented in Figure 3.13(e). The line de-
3.13(g) shows the result when line fitting is applied using the least squares approach. Rectangle shape fitting was then applied as presented in Figure 3.13(h). The final segmented household is shown in the Figure 3.13(i).
3.5
Summary
This chapter has introduced the Site A and Site B test set data and provided the necessary context to the preprocessing of the satellite image data so as to segment the desired household areas. The segmentation process was fully described. The results were two collections of im- ages. Note that each image has a family size (class label) associated with it. The distribution of these class labels was given in Table 3.1. Note that some initial experiments (not reported here) directed at evaluating the relationship between family size and household image size found that there was no correlation between the two. It was conjectured that this might be because poorer families lived in smaller households than richer families of the same size. Therefore, in the following three chapters three different techniques, whereby the satellite household image data can be classified are described and evaluated. The first technique considered is a graph-based technique which is presented in the following chapter.
Chapter 4
Population Estimation Mining using
Satellite Imagery: The Graph-Based
Approach
4.1
Introduction
This chapter considers the first of the three image representation approaches considered in this thesis, the colour histogram based and texture based representations are considered in Chapters 5 and 6 respectively. The idea promoted in this chapter is to capture the nature of each segmented household image, using a graph-based representation. In the case of the training data each graph represented household has a “family size” class label associated with it (see Table 3.1 presented in previous chapter). This training data can then be used to build a graph-based classifier that can be used to predict household sizes according to the nature of the proposed graph structure representation.
More specifically, in this chapter an image decomposition approach is considered whereby the individual households are represented using a quadtree (graph) decomposition; we refer to this as “fine segmentation” to distinguish it from the “coarse segmentation” used to identify households as described in Chapter 3. Once a set of households has been fine segmented the next stage of the data preparation phase is to translate the segmented pixel data into a form suitable for the application of a classifier. The translation needs to be conducted in such a way that all salient information is retained while at the same time ensuring that the representation is concise enough to allow for effective further processing. The fundamental idea here is to adopt a graph-based representation, more specifically a quadtree based representation (one per household). Quadtrees have been used extensively in the context of image processing (see for example [159]). However, the quadtree representation does not lend itself to ready incorpora- tion with respect to classification algorithms. To do this we propose applying subgraph mining to the quadtree data to identify frequently occurring patterns across the data that can be used as features in the context of a feature vector representation. The patterns of interest are thus
4.1
Figure 4.1: Schematic illustrating the Graph-Based Framework
A schematic of the graph-based approach is given in Figure 4.1. From the figure it can be seen that the graph-based approach for census mining consists of three processes: (a) image segmentation, (b) feature extraction and (c) classifier generation. The image segmentation process (the top rectangular box in the figure) was discussed in Chapter 3 and will thus not be considered further in this chapter. Once a set of individual segmented household has been identified, the next process, feature extraction, is used to translate the segmented pixel data into an appropriate form suitable for classifier generation (the third process within the framework). The classifier generation process is straight forward and requires little further consideration here.
The feature extraction process comprised a number of sub-processes (as shown in Figure 4.1). The fundamental idea underpinning the graph-based approach was to use a quadtree based representation (one per household). In this manner a set of subgraphs that frequently occur in the data could be identified which could then be used with respect to a feature vector representation of the form used by many classifier generation algorithms. The sub-processes
that make up the feature extraction process are: (i) quadtree decomposition, (ii) tree/graph representation, (iii) frequent subgraph mining, (iv) feature vector generation and (v) feature selection. Note that with respect to Figure 4.1 the approach presented is generic in nature and, as will be seen later in this thesis, similar approaches were used with respect to the colour histogram based and texture based approaches presented in Chapters 5 and 6 respectively.
The rest of this chapter is organised as follows. Quadtree decomposition is considered in Section 4.2, whilst Section 4.3 provides the detail of the tree/graph representation sub-process. The frequent subgraph mining and feature vector generation sub-processes are described in Section 4.4. Feature selection is then discussed in combination with classifier generation in Section 4.5. The evaluation of the proposed graph-based framework for population estimation mining using satellite imagery is then presented in Section 4.6. Finally, some further discussion and a summary are presented in Section 4.7 and 4.8 respectively.