Information Technology Reference
In-Depth Information
1
Full(30)
StarMiner(21)
ReliefF(21)
DTM(21)
0.95
0.9
0.85
0.8
0.75
Recall (%)
0.7
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
Fig. 7.5.
P&R graph built using the SegMRF test set represented by: 30 original
features, 21 selected by StARMiner, 21 selected by Relief-F and 21 selected by DTM
In Step 2, we measured the effectiveness of StARMiner algorithm in the task of
feature selection. To perform it, we also applied Relief-F to the training images.
The 21 most relevant features returned by Relief-F were also taken to compose
a feature vector. In addition, DTM was also applied to the training images and
the 21 most relevant features selected were also placed in a feature vector. The
StARMiner algorithm took 0.25 seconds to select the features, Relief-F took 0.72
seconds, and DTM took 0.85 seconds.
To build the Precision vs. Recall graphs, we considered four cases of feature
vectors used to represent the images: (a) using the 30 original features; (b) using
the 21 features selected by StARMiner; (c) using the 21 features selected by
Relief-F; (d) using the 21 features selected by DTM. Similarity queries were
executed over the test set and the P&R graphs were drawn. Figure 7.5 shows
the P&R graph obtained.
The graph in Figure 7.5 shows that the results obtained with 21 features are
quite better than the results gotten with all 30 features. Thus, although using
approximately 70% of the processing effort originally required, the precision of
content-based queries is improved (the computational effort of a similarity query
is proportional to the feature vector size).
To guarantee that we have selected the minimum set of relevant features that
maintain the precision results, we also executed the same
k
-nearest neighbor
