Image Processing Reference
In-Depth Information
• the execution times are larger when using radial cells, especially due to the trigonometric
calculus;
• over-increasing the cell size leads to performance drops, as the small textures are ignored.
The conclusion was that we will use square cells of 3 × 3 pixels and 8 pixels to compose the
local texture descriptor.
Subsequently, we have started experimenting with the rest of the descriptors. For the HOG
and SIFT algorithms we have used the authors' implementations. The tests involved the usage
of singular descriptors and combining all of them together; the final weights have been set as
it follows:
W 1 = 30% for the color space;
W 2 = 30% for the texture space;
W 3 = 40% for the local descriptors. These have been considered more representative than
the global descriptors.
The results are presented in Table 2 [ 16 ] .
Table 2
Experiments Involving All Descriptor Spaces
Descriptor Type
Results (%)
Color space 86
Texture space (LBP) 85
SIFT
82
HOG
85
All of the above
92
The results show that combining multiple types of descriptors from multiple search spaces
lead to performance improvements. On the above-mentioned data set, the results are prom-
ising and show an increase of over 5%. Also, the proposed architecture is able to correctly clas-
sify images obtained from document scans as well as regular images.
All the experiments conducted so far have been performed on document images which have
been correctly segmented. The next set of experiments have used images affected by different
types of noise, specific for the document analysis and recognition area.
As shown in Ref. [ 17 ] , the document scans can be affected by different alterations:
• extra characters or images being present in the scan, due to the paper transparency;
• page curvature introduces distortions;
• incorrect exposure to light;
• machine malfunctioning;
• additional ink spots, water stains, anchoring devices, other marks, etc.
Each of the above factors may affect the segmentation process and the overall classiication
results. Therefore we have used the CVSEG algorithm presented in Ref. [ 17 ] applied on a
mixed set of documents.
There are two areas which may be affected by the problems described above—the binariz-
ation stage or the segmentation stage. In what follows, we will describe how the binarization
results affect the segmentation stage and the overall performance.
 
Search WWH ::




Custom Search