7A-5-7
(a) (b)
Figure 13: (a) A pentagon imager(b) 10 magnitudes
of NFD of a pentagon
6 Text region detection
Traffic signs may have characters which indicate traf
fic meanings. UsuallyFsuch text traffic signs have an
uniform background color and an uniform foreground
color for text. Thus the texts have high contrast with
the background. A definition of texture energy can
be used for describing such contrast measurements. If
such text image is viewed in 3D spaceTthis measure
ment describe relative smoothness or roughness of the
flat plane.
Texture energy of a pixel can be defined on a local
window W centered at this pixel which is also call spa
tial variance widely applied in document image anal
ysis to separate text regions and image regions:
E =jrrj E (I(iJ)-i(iJ)) 2 (3)
(ij)ew'
where
= j E w.i»
and P is the number of pixels in the window W
Textness can be define by
textness =1 ——
1 + E
For a constant intensity window which all pixels
have the same valueTtexteness = OTand approaches
1 for large values of E (high contrast).
Text region detection. Texts in a traffic sign are
almost horizontal. A horizontal window (w x 1) can
be selected to compute the texture energy.
Sliding such window over the whole imageT a
textness image is obtainedT a local thresholding
method is used to segment this textness image. Be
cause text lines are horizontalThorizontal line-shaped
blobs with a certain size of area above are kept to be
masks to determine the text regions. To do soTtwo
measurement are computedrnamely eccentricity and
orientation.
The eccentricity is
_ (m 2 o - m 02 ) 2 + 4
(m 20 + m 02 ) 2
НегеГm pq is the moment of (pEi) orders. It ranges
from 0 to lTzero for a circular objectrand one for a
line-shaped object.
The orientation is
Ф — i arctan2(2mn,m 20 - m 0 2)
The horizontal line-shaped object will has an ap
proximate orientation of zero or ±n.
Aided by vertical edges. Directly usage of equa
tion (3) to compute textness will lead to too many
false alerts for text containing regions. For exampler
on borders of objects in an imageFonly one edge exists
in a local window may with high contrastrthe textness
of the pixel will be near to one. To disambiguate this
where the number of vertical edges in the local window
is too smallTthe texture energy is set to zero. Results
are shown in Figure 14 and Figure 15.
Connected component analysis. The above al
gorithm may not work on images taken in street or
in winterFwhich may contain frames of wall windows
or branches of trees. As we knowT characters in a
text lines are horizontally aligned and each line is sep
arated vertically. A further text extraction can be
done on the detected text regions based on connected
component analysis. Performed aft-егГа thresholding
method is used for further segmentation followed by
a connected component analysis. A minimal bound
ing rectangle which encloses each connected compo
nent is found. For a true text liner these rectangles
must approximately have the same size and align on
a horizontal line. This result can be used not only for
verification of text regionsTbut also in a traffic sign
recognition system. Results are shown in Figure 16.
Adaptive determination of local window size.
One problem with the above approach is determina
tion of the size of the local window. Small local win
dows cannot result in detection of text regions with
large fonts. To solve this ргоЫетГап adaptive text
location method is under development. One possible
method is to adaptively increase the size of the local