next up previous contents
Next: Future Work Up: report Previous: Segmentation using K-Means Algorithm   Contents

Experimental Results

In this Section, we present the results of the Gabor filter based script segmentation approach described above. The algorithm was tested on a few textures and multi-script documents. The textures used are the Brodatz textures. The algorithm works well for any placements of the textures within the image. The multi-script documents were obtained by combining different Indian scripts obtained from Indian web sites containing newspaper articles in Indian scripts. The following table shows the 5 top features for segmenting English and Hindi scripts:

Table 1: Features used for segmenting English and Hindi
Scale Orientation
3 0
6 162
2 24
6 12
6 174


The English script has a high response to the 0 degree Gabor filer because of the number of vertical segments in the script. The Hindi script is tilted a bit to its right and has a high response to the 162 and 174 degree filters. The following table shows the 5 top features for segmenting Tamil and English scripts:

Table 2: Features used for segmenting Tamil and English
Scale Orientation
5 150
6 24
6 156
4 12
1 0


The Tamil script is more curved than the English script and hence has a higher response to the 150, 156, 24 and 12 degree Gabor filters than the English script. The English script has higher response to the 0 degree filter. The following table shows the 5 top features for segmenting Gujarati and Hindi scripts:

Table 3: Features used for segmenting Gujarati and Hindi
Scale Orientation
4 90
6 24
6 66
2 60
2 24


The Hindi script has a high response to the 90 degree filter because of its headline feature. Note that this feature cannot be used for differentiating it from English which also has a high response to the 90 degree Gabor filter. The other features get selected because of the more rounded Gujarati script. The algorithm was tried on a few textures and the following results obtained:

\begin{figure*} \centerline{\epsfig{figure=2textures_final.eps,width=0.8\textwidth}} \end{figure*}

Figure 8: Result of segmenting two textures
\begin{figure*} \centerline{\epsfig{figure=one_inside_another_final.eps,width=0.8\textwidth}} \end{figure*}

The results obtained on some multi-script images are shown below:

Figure 9: Result of segmenting a document containing Tamil and English scripts
\begin{figure*} \centerline{\epsfig{figure=tamil_english_final.eps,width=0.9\textwidth}} \end{figure*}

Figure 10: Result of segmenting a document containing English and Hindi scripts
\begin{figure*} \centerline{\epsfig{figure=english_hindi_final.eps,width=0.9\textwidth}} \end{figure*}

Figure 11: Result of segmenting a document containing Gujarati and Hindi scripts
\begin{figure*} \centerline{\epsfig{figure=gujrati_hindi_final.eps,width=0.6\textwidth}} \end{figure*}

Figure 12: Result of segmenting a document containing English and Hindi scripts in a more complex layout
\begin{figure*} \centerline{\epsfig{figure=english_hindi_different_layout_final.eps,width=0.6\textwidth}} \end{figure*}

Figure 13: Result of segmenting a document containing Tamil, English and Hindi scripts
\begin{figure*} \centerline{\epsfig{figure=tamil_english_hindi_final.eps,width=0.9\textwidth}} \end{figure*}

next up previous contents
Next: Future Work Up: report Previous: Segmentation using K-Means Algorithm   Contents
2002-06-03