Segmentation of heterogeneous document images : an ... - Tel
Segmentation of heterogeneous document images : an ... - Tel
Segmentation of heterogeneous document images : an ... - Tel
You also want an ePaper? Increase the reach of your titles
YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.
(a) Five zones 1-5 (b) Projection pr<strong>of</strong>ile<br />
<strong>of</strong> zone 3<br />
(c) first derivative (d) initial, refined<br />
<strong>an</strong>d final regions<br />
Figure 2.13: Steps for locating text line separators in part <strong>of</strong> <strong>document</strong> image. [75]<br />
tel-00912566, version 1 - 2 Dec 2013<br />
less th<strong>an</strong> 2 ◦ . This method is successful provided that the free parameters are<br />
set correctly. Furthermore, because the algorithm keeps track <strong>of</strong> the domin<strong>an</strong>t<br />
direction, the whole <strong>document</strong> must have text lines with roughly the same direction.<br />
The method published in [61] is one more method based on Hough tr<strong>an</strong>sform.<br />
In the first step, authors apply a Hough tr<strong>an</strong>sform to each connected<br />
component namely the h<strong>an</strong>dwritten words to find the direction <strong>of</strong> a component.<br />
Then the algorithm searches for the nearest neighbors <strong>of</strong> each component in four<br />
principal directions. Once the neighbors are found, a weighted directed graph<br />
is built by connecting each component to its neighbors with a weighted edge<br />
proportional to the geometric dist<strong>an</strong>ce between components. Finally, to form<br />
text lines, the algorithm removes top to bottom edges based on thresholding<br />
the length <strong>of</strong> edges.<br />
Texture based methods<br />
Any method that is based on some kind <strong>of</strong> filtering, shall it be Gabor, Wavelet,<br />
Gaussi<strong>an</strong> or just the averaging operator c<strong>an</strong> fit into this category.<br />
The first method that we review is for text line segmentation from freestyle<br />
script-independent h<strong>an</strong>dwritten or printed <strong>document</strong>s. Y. Li et al. first have<br />
published their preliminary results for this method in [52] <strong>an</strong>d later in [53]. For<br />
this method it is assumed that text lines have a horizontally elongated shape,<br />
but still a variation <strong>of</strong> ±10 ◦ is allowed. The method estimates a probability density<br />
function based by convolving the image with a non-parametric <strong>an</strong>isotropic<br />
Gaussi<strong>an</strong> kernel. The initial estimates <strong>of</strong> the text line boundaries are computed<br />
by thresholding this density function map <strong>an</strong>d then a level set method evolves<br />
from the initial estimations to obtain the final text line boundaries.<br />
Another method is proposed in [29]. Du et al. propose a script-independent<br />
method for segmentation <strong>of</strong> h<strong>an</strong>dwritten text lines based on a piecewise ap-<br />
31