Mobile Landmark Recognition - Multimedia Database Retrieval: Technology and Applications - page 140

5.5.3 Recognition Function

In a landmark recognition system (e.g., [ 346 ]), the image database contains images
I 1 ,

. Thus,
these available class labels can be assigned to a new landmark image queried by
a user. For this recognition task, the system obtains the matching image using
D d (

I 2 ,..., I N , each of which belong to one of the c -classes, { C 1 , C 2 ,..., C c }

N and the best matching image is selected. Its class label
is then assigned to the given query I q .
In the current work, the class assignment to the query is performed differently,
by using the ranking list. The top- R retrieved images are firstly obtained, i.e.,
Query

I x , I q ) , ∀ x = 1 , 2 ,...,

I q )= { I q ) ≤ I q ) }

. The class labels of these
images are also retrieved. Since the modified query A q is expected to be more
effective than the original query, the retrieved image set will contain more relevant
images, i.e., higher probability of relevant images in the top- R retrieved images. The
probability can be measured by:

( A 1 , A 2 ,..., A R | D d ( A i , D d ( A R ,

} )= |

C c | R
R

( |{ , ,...,

P C c A 1 A 2 A R (5.32)

where C c | R is the number of occurrences of the c -th class label, C c of the retrieved

images in the top- R best matches. With this definition, the class label assigned to
the query is obtained by the class labels that occur the most frequently in the top
matches, i.e.,

|

c ∗ = arg max

c

P ( C c |{ A 1 , A 2 ,..., A R } ) (5.33)

The query image will be assigned to the class c ∗ of the landmark images in the
dataset.

5.6 Experimental Result

In the experiment, two datasets of landmark images were constructed. The first
database is the Singapore landmark dataset containing 50 categories with 4,060
images in total. The second database is the World landmark dataset containing
72 categories with 8,847 images in total. Each category is associated with the
landmark, and includes multiple images of each landmark. These images were
originally collected from the internet. Sample images from the Singapore and World
landmark databases are shown in Fig. 5.4 . Images were divided into two subsets:
training and testing for demonstrating the recognition performance of the system.
Figure 5.5 shows samples of training and testing images from the Art Science
Museum category in the Singapore landmark database. These images were taken
at different viewpoints, and their apparentness are varied according to the camera
viewpoints.