Kernel-Based Adaptive Image Retrieval Methods - Multimedia Database Retrieval: Technology and Applications - page 28

Table 2.5 Comparison of adaptive retrieval methods

Method Learning algorithm

RBF-1 RBF center model 1 ( ʱ = 1 . 2 , ʱ = 0 . 08 ) R N

RBF width model 1 ( ʷ = 12 )

RBF-2 RBF center model 2 ( ʱ R = 1 . 2)

BRF width model 1

( ʷ = 12 )

QAM [ 12 , 14 , 18 , 36 ,
51 , 103 ]

Query modification in Eq. ( 2.8 ), Cosine similarity metric in Eq. ( 2.6 )
( ʱ = 1 , ʲ = 4 , ʳ = 0 . 8 )

MAM [ 17 , 20 , 21 ] City-block distance is used for similarity metric. The feature weighting

is obtained by the standard deviation criterion

Table 2.6 Average
precision (%)

Method Iter. 0 Iter. 1 Iter. 2 Iter. 3

MAM 76.70 80.43 80.71 80.74

RBF-1 76.70 85.02 86.90 88.12

RBF-2 76.70 85.32 86.80 87.37

QAM 67.10 75.12 76.42 76.95

lowest performance gains and converges at about 62 % precision. It is observed
that the learning capability of RBF is more robust than that of OPT-RF, not only
in retrieval capability, but also learning speed. As presented in Table 2.9 , results
after one round of the RBF method are similar to results after three rounds of the
OPT-RF method. This quick learning is highly desirable, since the user workload
can be minimized. This robustness follows from imposing nonlinear discriminant
capability in combination with positive and negative learning strategies.
Typical retrieval sessions are shown in Fig. 2.2 , for the Yacht query. Figure 2.2 a
show the 16 best-matches images before applying any feedback, with query image
display in the top-left corner. It was observed that some retrieved images were
similar to the query image in terms of color composition. In this set, three retrieved
images were marked as relevant subjects to the ground truth classes. Figure 2.2 b
shows the improvement in retrieval after three rounds of using the RBF learning
method. This is superior to the results obtained by MAM (cf. Fig. 2.2 c) and OPT-
RF (cf. Fig. 2.2 d). This query may be regarded as a “hard” query, which requires
a high degree of nonlinear discrimination analysis. There are some quires that
are relatively easier to retrieve, which are shown in Fig. 2.3 . Those queries have
prominent features, such as a shape in the Rose query, and a combination of texture
and color in the Polo query. In each case, it is observed that the MAM and OPT-
RF methods show better performance than in the previous results. In these cases,
however, the retrieval results obtained by RBF approached 100 % precision.