Image Processing Reference
In-Depth Information
4.3 Edge Features
Canny edge detection algorithm [ 14 ] is one of the most reliable algorithms for edge detection.
Our results show that for most cases, the shapes of crystals are kept intact in the resulting edge
image. An edge image can contain many edges which may or may not be part of the crystals.
To analyze the shape and other edge related features, we link the edges to form graphs. We
used the MATLAB procedure by Kovesi [ 17 ] to perform this operation. The input to this step
is a binary edge image. First, isolated pixels are removed from the input edge image. Next, the
information of start and end points of the edges, endings and junctions are determined. From
every end point, we track points along an edge until an end point or junction is encountered,
and label the image pixels. The result of edge linking for the Canny edge image in Figure 4 (b)
is shown in Figure 4 (c).
FIGURE 4 Edge detection and edge feature extraction: (a) original image, (b) Canny edge
image, (c) edge linking, (d) line fitting, (e) edge cleaning, and (f) image with cyclic graphs or
edges forming line normals.
Two points (vertices) connected by a line is an edge. Likewise, connected edges form a
graph. Here, we use the terms edges and lines interchangeably. The number of graphs, the
number of edges in a graph, the length of edges, angle between the edges, etc. are good fea-
tures to distinguish different types of crystals. To extract these features, we do some prepro-
cessing on the Canny edge image. Due to the problem with focusing, many edges can be
formed. To reduce the number of edges and to link the edges together, line iting is done. In
this step, edges within certain deviation from a line are connected to form a single edge [ 17 ] .
The result from line iting is shown in Figure 4 (d). Here, the margin of three pixels is used
as the maximum allowable deviation. From the figures, we can observe that after line fitting
the number of edges is reduced and the shapes resemble to the shapes of the crystals. Like-
wise, isolated edges and edges that are shorter than a minimum length are removed. The res-
ult from removing the unnecessary edges is shown in Figure 4 (e). At the end of edge linking
procedure, we extract the eight edge related features listed in Table 2 . The lengths of edges are
calculated using Euclidean distance measure. Likewise, the angle between the edges is used to
determine if the edges (lines) are normal to each other. We consider two lines to be normals if
the angle θ between them lies between 60 and 120 (i.e., 60 ≤ θ ≤ 120).

Search WWH ::

Custom Search