21.01.2014 Views

New method for feature extraction based on fractal behavior - IDRBT

New method for feature extraction based on fractal behavior - IDRBT

New method for feature extraction based on fractal behavior - IDRBT

SHOW MORE
SHOW LESS

You also want an ePaper? Increase the reach of your titles

YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.

Y.Y. Tang et al. / Pattern Recogniti<strong>on</strong> 35 (2002) 1071–1081 1075<br />

Fig. 2. Diagram of <str<strong>on</strong>g>feature</str<strong>on</strong>g> <str<strong>on</strong>g>extracti<strong>on</strong></str<strong>on</strong>g> by wavelet sub-patterns and divider dimensi<strong>on</strong>s <str<strong>on</strong>g>for</str<strong>on</strong>g> the Chinese character “Heart”.<br />

Thus, the expressi<strong>on</strong>s <str<strong>on</strong>g>for</str<strong>on</strong>g> the wavelet trans<str<strong>on</strong>g>for</str<strong>on</strong>g>mati<strong>on</strong> of<br />

V 0 can be written as follows:<br />

5∑<br />

5∑<br />

c j+1;m = h k c j;k+2m ; d j+1;m = g k c j;k+2m (15)<br />

k=0<br />

k=0<br />

where m =0; 1;:::;N − 1.<br />

Hence, the wavelet trans<str<strong>on</strong>g>for</str<strong>on</strong>g>mati<strong>on</strong> sub-patterns of V 0<br />

obtainedusing Eq. (16) will become<br />

V 1 = {c j+1;0 ;c j+1;1 ;:::;c j+1;N−1 };<br />

W 1 = {d j+1;0 ;d j+1;1 ;:::;d j+1;N−1 }: (16)<br />

Step 3. Computati<strong>on</strong> of divider dimensi<strong>on</strong>s <str<strong>on</strong>g>for</str<strong>on</strong>g><br />

wavelet sub-patterns: Since the divider dimensi<strong>on</strong>s of<br />

n<strong>on</strong>-self-intersecting curves are asymptotic values, we<br />

can derive their approximati<strong>on</strong>s <str<strong>on</strong>g>based</str<strong>on</strong>g> <strong>on</strong> the following<br />

expressi<strong>on</strong>:<br />

log M (C)<br />

−log ;<br />

when is set to be small enough. M (C) is the compass<br />

dimensi<strong>on</strong> computed with a ruler of size al<strong>on</strong>g the<br />

curve C.<br />

In our experiments, we per<str<strong>on</strong>g>for</str<strong>on</strong>g>medthree c<strong>on</strong>secutive<br />

wavelet trans<str<strong>on</strong>g>for</str<strong>on</strong>g>mati<strong>on</strong>s <str<strong>on</strong>g>for</str<strong>on</strong>g> each <strong>on</strong>e-dimensi<strong>on</strong>al<br />

pattern. Hence, the <strong>on</strong>e-dimensi<strong>on</strong>al pattern, such as<br />

the central projecti<strong>on</strong> of the Chinese character “Heart”<br />

(Fig. 2), will yieldthree n<strong>on</strong>-self-intersecting curves;<br />

namely: V 0 ;V 1 ;W 1 . By comparing these two frequency<br />

spectra, we can ndthat V 0 c<strong>on</strong>tains all frequency comp<strong>on</strong>ents<br />

of f(x), while V 1 c<strong>on</strong>tains <strong>on</strong>ly <strong>on</strong>e half of the<br />

frequency comp<strong>on</strong>ents of f(x). The reas<strong>on</strong> is that V 0<br />

is divided into two parts: V 0 = V 1 ⊕ W 1 . As <str<strong>on</strong>g>for</str<strong>on</strong>g> the rst<br />

part, V 1 , <strong>on</strong>ly the low-frequency comp<strong>on</strong>ents of V 0 are<br />

kept, andthe high-frequency comp<strong>on</strong>ents are lost. In<br />

additi<strong>on</strong>, <strong>on</strong>ly the high-frequency comp<strong>on</strong>ents of V 0 are<br />

kept, andthe low-frequency comp<strong>on</strong>ents are lost in W 1 .<br />

For each of the three curves, we further computedits<br />

divider dimensi<strong>on</strong>, and there<str<strong>on</strong>g>for</str<strong>on</strong>g>e, related each pattern<br />

with a <str<strong>on</strong>g>feature</str<strong>on</strong>g> vector. These <str<strong>on</strong>g>feature</str<strong>on</strong>g> vectors are presented<br />

as follows:<br />

D vo =1:03911917; D V1 =1:03940148;<br />

D W1 =1:60938780:<br />

These <str<strong>on</strong>g>feature</str<strong>on</strong>g> vectors are selectedto be usedas the<br />

<strong>fractal</strong> <str<strong>on</strong>g>feature</str<strong>on</strong>g> <str<strong>on</strong>g>for</str<strong>on</strong>g> extracting the <str<strong>on</strong>g>feature</str<strong>on</strong>g>s of object. More<br />

details about the central projecti<strong>on</strong> trans<str<strong>on</strong>g>for</str<strong>on</strong>g>mati<strong>on</strong> and<br />

the orth<strong>on</strong>ormal wavelet decompositi<strong>on</strong> can be found in<br />

Ref. [14].<br />

In our experiments, we have testedour new approach<br />

<str<strong>on</strong>g>for</str<strong>on</strong>g> two types of images: 52 upper andlower case printed<br />

alphabets and3000 Chinese characters. We classify<br />

these <str<strong>on</strong>g>feature</str<strong>on</strong>g>s andper<str<strong>on</strong>g>for</str<strong>on</strong>g>m the recogniti<strong>on</strong> of dierent<br />

types of patterns by using the least distance classicati<strong>on</strong><br />

<str<strong>on</strong>g>method</str<strong>on</strong>g>. For each type of the patterns, the center

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!