A content-based image retrieval approach based on document queries - Emerging Trends in Image Processing, Computer Vision, and Pattern Recognition - page 266

There is also a small performance drop when processing rotated images but usually this
problem is solved by the rotation invariance characteristic in the descriptors.
In what regards the segmentation experiments, we have used the CVSEG algorithm [ 17 ] . We
have also found some problems in what regards processing the images at this stage. Usually
they are caused by:
• incorrect binarization results—this may lead to segmentation results which include too
much or too litle from the original image;
• usage of different font types—some of the larger fonts may get included in the result;
• page layout, borders, page rotation, objects which did not belong in the original document,
etc.
The results are described in Table 4 .

Table 4
Segmentation Experiments

Segmentation Problem Overall Results (%)

Binarization 74.2

Font type 91.0

Page layout 91.1

Again, the binarization results have an important impact on the overall performance. As
mentioned before, the segmentation module cannot extract the correct image from a document
which has not been correctly binarized, because in many cases the original drawings/images
have been masked by the large black areas. Subsequently, the classification module cannot ex-
tract the appropriate descriptors and fails to correctly classify the image.
There are some small performance drops in the remaining two areas as well but usually,
even the segmentation results contain erroneous areas (page borders or extra text) they have
litle or no impact at all in the classification stage. Most of the problems in this area have been
caused by the incomplete images originating in the segmentation results.
The last experiments have been conducted on mixed images, with problems in both the bin-
arization and segmentation stages, or no problems at all, originating from the ICDAR dataset
and from the Ubuntu documentation. The combined results are described in Table 5 .

Table 5
Overall Classification Performance Fluctuations

Images Average Results (%) Performance Drop (%)

Binarization problems 85.13 6.87

Segmentation problems 85.43 6.57

Mixed images 89.58 2.42

As expected, the overall performance dropped especially because of the results obtained in
the binarization stage, which impact the remaining stages as well. The difference however is