A content-based image retrieval approach based on document queries - Emerging Trends in Image Processing, Computer Vision, and Pattern Recognition - page 260

(6)
• the whole image in RGB coordinates;
• the RGB histogram, with 256 bins.
Each of the four sets of characteristics is used as an input for a standard feed forward/back
propagation neural network. The neural networks' outputs are then collected by a simpliied
weighted majority voting module, as it can be observed in Figure 2 .
FIGURE 2 Color space classification.
The weighted module works according to the below algorithm:
• let n be the number of accepted classes and k be the number of classifiers;•
• each neural network will produce on the final layer a vector C x = { c 1 , c 2 , …, c n }, where
1 ≤ x ≤ k ;
• the weight associated to the output layer will be W = { w 1 , w 1 , …, w k };
• the weighted result will be provided by the wiCi sum, as specified below, where R ε [1, n ],
max( C ) represents the maximum value obtained for a certain class, and idx represents the
position of this class in the final vector
(7)
In the texture space area we have chosen an approach based on local binary patern
descriptors, mainly because of their invariant properties for color or rotation.
For the local descriptors we have chosen two sets of characteristics, based on SIFT and his-
togram of oriented gradients (HOG). Traditionally, the HOG descriptors are used in order to
train an SVM classifier, but since we are dealing with a multiple classification problem, we
have used neural networks in the learning stage for both descriptors. The two types of local
descriptors produced similar results during the tests, therefore the combined classifier for SIFT
and HOG uses equal weights of 50%.
The final classifier includes an additional weighted majority voting module, as shown in
Figure 3 .