Hierarchical representation of features is not a new approach in CBIR. Various researchers used the concept of representing objects in hierarchical form. Xu et al. [150] propose a system that employs a hierarchical content tree data structure for every image in a database. Similarity of content is computed by searching this representation in a top-down fashion, by first matching composite nodes (objects) if they are already formed, to reduce search time, and thereafter, matching combinations of elementary nodes if match has not been established at a higher level.
Features are hierarchically divided into four layers, namelylocal layer,crack-network layer,
approach taken can be described as a structural scaling approach, which views objects-of- interest at hierarchically-separated yet related structural levels. The layers are organised in a data structure, so as to allow direct access to their components and to prepare a basis for further manipulation of their embedded data.
1st layer (local layer) 2nd layer (crack-network layer) 3rd layer (global layer) 4th layer (image layer)
Figure 6.1: Hierarchy of features.
The first layer (local layer) is more like a “hidden layer”, since it is not used explicitly as an object descriptor. However, its importance is quite significant, due to the fact that it functions as a “root” to enable the higher level features to “grow”. This first layer of interest concentrates on fine entities, which involve line segments. The primary assumption is that a crack-network consists of multiple line segments. However, this assumption can be relaxed with crack-networks that do not contain more than a single line. In this case, there are three layers in the representation, i.e. the local/crack-network layer, the global layer and theimage layer.
The second layer (crack-network layer) is formed from combinations of the first layer. Every node in the first layer has a corresponding significance value regarding the formation of a higher level. Thesignificance measure is explained in Section6.3.1. The object created at this layer is the entity of a crack pattern which is denoted as a crack-network.
combination of entities from the crack-network layer. The process of combining these entities has already been explained thoroughly in Chapter 5. It is important to note that not all features are computed using weightings or significance measures. Some features are calculated by straightforward assessment of the structural characteristics of a particular layer in most cases, the global layer.
The upper-most layer in the hierarchy is the image layer, which comprises every single entity within an analysed image.
6.2.1 The Basic Features
From a global point of view, a single crack-network holds information regarding the number of local entities detected by thecrack following routine, where these include the number of nodes and line segments. These entities are accumulated as the crack contour is “followed”. Basic features are also computed for use in higher level feature extraction processes.
The perimeter is computed as the length of a chain [50,151]. The formula for the perimeter is
P = ne+
√
2no (6.1)
where ne is the number of even chain elements and no the number of odd chain elements
(based on chain-code connectivity as in Figure1.2(b)) . The total length of a crack-network length is calculated using Equation6.1and its significance will be explained in Section6.3.
One of the most important items of information that needs to be captured by the crack following routine is line orientation. For this purpose, chain-codes are accumulated in a histogram with 8 bins, where each bin represents a chain-code. This histogram is called the orientation histogram. It roughly indicates the orientation spread of a particular crack- network. Thus, globally, it can be determined whether there are any dominant directions or whether the directions are equally spread. Figure 6.2 shows an example of orientation histograms for line segments.
Each crack-network consists of line segments and these line segments are connected by nodes. The two components (line segments and nodes) are considered as the local entities of a crack-network. The important feature of a node is its location in a crack-network. As a whole, this generally informs how concentrated or sparsely distributed nodes are in a crack-network.
(a) 1 2 3 4 5 6 7 8 0 50 100 150 200 250 300 350 Chain code Accumulation (b)
Figure 6.2: A crack contour (a) with its orientation histogram (b).
Line segments obtain almost the same type of information compared to the crack-network layer, except that they capture statistical data on a line-to-line basis. This means that each line segment has its own features. For every line segment, the edge points are marked and the length is recorded using Equation6.1. Orientation histograms are also constructed for each line segment.
A complete crack-network data structure allows straightforward manipulation of crack pat- tern entities. This simple framework is useful for the later stages of extracting high-level features.