Kernel-Based Adaptive Image Retrieval Methods - Multimedia Database Retrieval: Technology and Applications - page 30

Table 2.7 Database and feature extraction methods

Item Description

Corel database
[ 73 ]

The database contains 40,000 real-life images divided into 400 classes.
Every class has 100 images

Color histogram
[ 38 ]andcolor
moments

The first descriptor is a 48-bin color histogram in HSV color space.
The second descriptor is a nine-dimensional vector, conducted from the
mean, standard deviation, and skew of the three RGB-color channels

Gabor wavelet
descriptor [ 91 ]

The 48-dimensional descriptor contains the mean and standard devia-
tions of the Gabor wavelet coefficients from the filtering in four scales
and six orientations

Fourier
descriptor [ 37 ]

The nine-dimensional descriptor contains the Fast Fourier transform
coefficients (at low frequency) of the edge information of an input image

Table 2.8 Comparison of adaptive retrieval methods

Method Learning algorithm

RBF RBF center model 2, RBF width model 2

OPT-RF [ 26 ] Optimum query adaptation model Eq. ( 2.14 ), optimum weighting metric

Eq. ( 2.15 ), Mahalanobis distance Eq. ( 2.11 ) as similarity function

MAM [ 17 , 20 , 21 ] Mahalanobis distance Eq. ( 2.11 ) as similarity function, weight parameters

Eq. ( 2.10 ) are obtained by the standard deviation criterion

Table 2.9 Average precision rate (%) obtained by retrieving 35 queries selected
from different categories, using the Corel database (columns 2-5)
Method Iter. 0 Iter. 1 Iter. 2 Iter. 3 CPU time (Sec./Iter.)
RBF 44.82 79.82 88.75 91.76 2.34
MAM 44.82 60.18 61.61 61.96 1.26
OPT-RF 44.82 72.14 79.64 80.84 1.27
Non-adaptive method 44.82 - - - 0.90
Average CPU time obtained by retrieving a single query, not including the time to
display the retrieved images, measured from a 1.8 GHz Pentium IV processor and
a MATLAB implementation

own local characteristics. The difficulty in characterizing image relevancy, then,
is identifying the local context associated with each of the sub-classes within
the class plane. Human beings utilize multiple types of modeling information to
acquire and develop their understanding about image similarity. To obtain more
accurate, robust, and natural characterizations, a computer must generate a fuller
definition of what humans regard as significant features. Through user feedback,
computers do acquire knowledge of novel features which are significant but have not
been explicitly specified in the training data. This implicit information constitutes
subclasses within the query, permitting better generalization. In this case, a mixture
of Gaussian models is used, via the RBF network, to represent multiple types of
model information for the recognition and presentation of images by machines.