14.01.2014 Views

Segmentation of heterogeneous document images : an ... - Tel

Segmentation of heterogeneous document images : an ... - Tel

Segmentation of heterogeneous document images : an ... - Tel

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.

(a) Document image (b) Small Gabor kernel (c) Large Gabor kernel<br />

tel-00912566, version 1 - 2 Dec 2013<br />

Figure 4.7: Results <strong>of</strong> applying two Gabor filters with different kernel sizes to a<br />

<strong>document</strong> image. This clearly shows the ability <strong>of</strong> Gabor filters to capture text lines<br />

<strong>of</strong> different font sizes. The result in b belongs to a Gabor filter with medium kernel<br />

size. As the size <strong>of</strong> the Gabor filter gets larger in c, it reveals larger text lines on the<br />

page.<br />

site. To reduce the effect <strong>of</strong> scale, me<strong>an</strong> <strong>an</strong>d vari<strong>an</strong>ce are not computed using<br />

the same height <strong>an</strong>d width as the site but by using a patch centered on the<br />

site but with a size proportional to the local height <strong>of</strong> text components (which<br />

comes from the height map) at that site.<br />

Some global feature functions are used that do not depend on <strong>an</strong>y observation.<br />

These functions are noted below. In these function y c refers to the label<br />

<strong>of</strong> the current site at the center <strong>an</strong>d y t , y l <strong>an</strong>d y tl refer to labels on the top,<br />

left <strong>an</strong>d top-left <strong>of</strong> the site, respectively. Labels may be 1 for text <strong>an</strong>d 0 for<br />

non-textual sites. Note that these functions are separate independent feature<br />

functions that each takes its own weight while training. Thus, they c<strong>an</strong>not be<br />

merged into a single function.<br />

f = [y c = y l ]<br />

f = [y c = y t ]<br />

f = [y c = 0] × [y l = 0]<br />

f = [y c = 0] × [y l = 1]<br />

f = [y c = 1] × [y l = 0]<br />

f = [y c = 1] × [y l = 1]<br />

f = [y c = 0] × [y t = 0]<br />

f = [y c = 0] × [y t = 1]<br />

f = [y c = 1] × [y t = 0]<br />

f = [y c = 1] × [y t = 1]<br />

64

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

Saved successfully!

Ooh no, something went wrong!