Self-adaptation in Image and Video Retrieval - Multimedia Database Retrieval: Technology and Applications - page 75

Table 3.2 Average retrieval
rate (AVR) of 116 query
images on DB1, obtained by
pseudo-RF learning

Method Initial Iter. 1 Iter. 2 Iter. 3 Iter. 4

Retrieval based on WM descriptor
ARBFN 58.78 69.02 72.85 76.24 77.21

Single-RBF 58.78 66.32 68.80 70.04 71.87

QAM 53.88 57.11 59.00 60.45 60.78

Retrieval based on NHI descriptor
ARBFN 63.42 71.66 75.22 75.86 76.51
Sigle-RBF 63.42 70.31 72.74 73.11 73.06
QAM 60.35 67.89 71.07 72.63 72.79
The initial AVR results were obtained by
Euclidean metric for ARBFN and single-RBF, and
by Cosine measure for QAM

3.3.3.1 Pseudo-RF Result

Table 3.2 provides numerical results illustrating the performance of the pseudo-RF
method on DB1 database. In all cases, relevance judgment was based on the ground
truth. Three adaptive retrieval methods: adaptive radial basis function network
(ARBFN), the single-radial basis function (RBF) method, and the query adaptation
method (QAM) were tested (these methods are discussed in Chap. 2 ). For each
learning method, the top 20 ranked images were utilized as training data. These
samples were input to the SOTM algorithm for pseudo labeling. The output of the
unsupervised network was in turn used as the supervisor for RF learning to update
learning parameters and to obtain a new set of retrievals.
The use of pseudo-RF learning for automatic retrieval resulted in a significant
improvement in retrieval performance over that of the non-adaptive technique. For
the automatic ARBFN, 18.4 % improvement in average precision was achieved
through four iterations of pseudo-RF, whereas the automatic single-RBF provided
a 13 % improvement. These retrievals used the WM descriptor. The results for each
learning method, with the MHI descriptor, show the same trend.
Figure 3.7 provides an example of a retrieval session performed by the automatic
ARBFN learning method, using the WM descriptor. Figure 3.7 a shows retrieval
results without learning, and Fig. 3.7 b shows the results after automatic learning.
The improvement provided by the automatic retrieval method is apparent.

3.3.3.2 Retrieval Results of Semiautomatic Retrieval

In order to verify the performance of the unsupervised learning of the pseudo-
RF retrieval system, its performance was compared with that of the traditional RF
method. The retrieval system was allowed to interact with the user to perform the
retrieval task, and the results obtained are provided in Table 3.3 . It was observed that
user interaction gave better performance: 3.34-6.79 % improvement was seen after
one iteration, and 3.66-4.74 % after four iterations. However, it should be taken into