Image Processing Reference
In-Depth Information
Table 1
Color Space Experiments
Combined Descriptors Results (%)
RGB + l1l2l3 82
RGB + c1c2c3 84
RGB + RGB histograms 69
As we can observe, the presence of the c1c2c3 and l1l2l3 color spaces produces an improve-
ment of over 10%, leading to the below conclusions:
• the experiments conducted on real world images confirm the necessity of additional color
space descriptors;
• c1c2c3 produces the most solid performance boost. After analyzing the images in the data
set, we have observed the presence of many pictures with shady areas, which shows that
this color space is adequate for these conditions. Figure 5 shows the effects of the c1c2c3
normalization on an image containing highlight and shadow areas;
FIGURE 5 c1c2c3 normalization.
• introducing the RGB histograms as a global descriptor actually produced a performance
drop, as two different images may have very similar color histograms, yet a very different
content. This was mainly caused by the surrounding conditions in which the picture of a
certain object was taken. Also, the presence of the shadow areas affects the histograms and
implicitly the classification result.
After experimenting with various color space weight values we have chosen the below val-
• the RGB histograms weights have been set to w 4 = 10%, which improved the overall per-
formance. We have kept this set of descriptors for situations where the color plays a more
important role in the classification process. In this case, the user will be able to adjust this
value accordingly;
• the rest of the weights have been set to w [1:3] = 30% (for the RGB/c1c2c3/l1l2l3 descriptors).
This lead to a combined overall performance of 86%. As we mentioned before, the UI ofers
the user the possibility of manually adjusting the global color relevance (associated to the
RGB histograms) in the final result. As an example, the user can choose a combination like
w 1 = 20%, w 2 = 20%, w 3 = 20%, and w 4 = 40%.
The next set of experiments was conducted in the texture space, with the LBP descriptor.
The main problems in this area were related to choosing the cell shape and size, along with
the number of pixels which compose the final descriptor. After a series of tests we concluded
the below:
• choosing radial cells over square cells produces an overall performance increase of over
10%, going over 95%;
Search WWH ::

Custom Search