O'Reilly logo

Logo Recognition by Dan Chen, Lizhe Wang, Jingying Chen

Stay ahead with the world's most comprehensive technology and business learning platform.

With Safari, you learn the way you learn best. Get unlimited access to videos, live online training, learning paths, books, tutorials, and more.

Start Free Trial

No credit card required

4
System overview
CONTENTS
4.1 Preprocessing .............................................................. 53
4.2 Polygonal approximation .................................................. 55
4.3 Indexing ................................................................... 55
4.4 Matching .................................................................. 58
The logo recognition system comprises four modules, i.e., preprocessing, polyg-
onal approximation, indexing and matching modules. The preprocessing mod-
ule operates on raw inputs of the logo images. Outputs of matches are dis-
similarity distances between models and test images. The system flowchart is
illustrated in Figure 4.1, where each of the processes is described briefly in
the following. The processes with are the focus of this work.
4.1 Preprocessing
Preprocessing is the first phase of the system chart shown in Figure 4.1. Af-
ter scanning, the edge detection and thinning process will be performed to
segment the logo. Then contour extraction will be employed to generate se-
quential lists of image pixel locations. The output of this phase will be sent
to the polygonal approximation module.
Edges of intensity images are usually used as important features. Edge
detection is to detect discontinuities in the image intensity. It has been stud-
ied most extensively, and many reliable algorithms have been proposed and
implemented [29, 56, 186, 99, 136, 149, 150, 154]. Heath et al. investigated the
performance of different edge detectors. They compared edge detectors based
on experimental psychology and statistics, in which humans rated the output
of low level vision algorithms. One of their clear results is “No one single edge
detector was best overall; for any given image it is difficult to predict which
edge detector will be best” [79]. In this study, the proposed logo recognition
method does not rely on any specific edge detector. Any edge detector can be
used, and the implementation from Nevatia and Babu [150] is adopted in this
study. After edge detection, a thinning process is used to reduce thick edges
53
54 Logo Recognition: Theory and Practice
FIGURE 4.1: The system flowchart of logo recognition. The processes with
are the focus of this work.
to chains of single pixels that can be easily traversed. Examples of the edge
image and thinned image are shown in Figure 4.2.
System overview 55
(a) (b) (c)
FIGURE 4.2: An illustration of segmentation: (a) intensity image, (b) edge
image, (c) thinned image.
Contour is a compact way to represent the shape of an image. It stores
sequential lists of pixel locations which can be represented by integer coordi-
nates. Given a sequence of integer coordinate points p
1
(x
1
,y
1
), p
2
(x
2
,y
2
),
..., p
n
(x
n
,y
n
), where p
i
is connected to p
(i+1)
,1in.Wehave|x
i
x
i+1
|
and |y
i
y
i+1
| both less than or equal to one but not both equal to zero. Con-
tour extraction is very important because the quality of the resulting contours
may affect the following processes.
4.2 Polygonal approximation
The polygonal approximation module is responsible for extracting feature
points from the extracted contours of the shape. This part is important for
the logo recognition system. In addition, the shape of an object can be rep-
resented in a more compact way. Computation time can thus be saved. The
key requirement of this part is that the points extracted must represent the
shapes faithfully. Missing points and spurious points should be avoided as
much as possible. In order to tackle these problems, a new feature point de-
tection method based on robust shape feature is proposed in this study. More
details can be found in the next chapter.
After feature point detection, every line segment that connects two con-
secutive points on the contour is used to represent the image. These line
segments are used as matching primitives here, since they possess good prop-
erties, i.e., (1) line segments are easier to obtain from digital images compared
to nonlinear features such as curves, (2) they represent higher level structure
as compared to contour points, (3) line segments occupy less memory than
contour point representation, (4) it is difficult to represent a curve if there are
gaps or missing/shifted feature points; line representation, however, has no
such problem. One example of the line segment map is shown in Figure 4.3.

With Safari, you learn the way you learn best. Get unlimited access to videos, live online training, learning paths, books, interactive tutorials, and more.

Start Free Trial

No credit card required