Digital Signal Processing Reference
In-Depth Information
5
Experimental Results and Analysis
In this section, we conduct image retrieval experiments on the proposed techniques
and system. Corel Photo Gallery is used as the image database, which contains 1000
JPEG images with size of 384x256 or 256x384 pixels. These images are composed of
10 semantic categories, such as Africa, buildings, horses, elephants, dinosaurs, flow-
ers, buses, beach, mountains, and food. Each category includes 100 images. 6 images
are selected randomly as query images from each of 5 image semantic categories and
hence, a total of 30 query images are taken in our experiments. For each query, we
select the top
k
results from the query results to compute precision, i.e.
p
.
p
is
defined as the proportion of the relevant images retrieved in the top
k
retrieved im-
ages according to the similar distance. The average precision (
p
) of 6 retrieval re-
sults are regarded as retrieved precision of each semantic category. Three methods,
namely traditional color histogram retrieval (TCHR),weighted color based color con-
trast histogram retrieval (WCHR) an
d
combining weighted color histogram and cohe-
sion (WCHCR), are compared with
p
when
k
.The comparison can be
shown in the Table 1. We can see that the method proposed (WCHCR) in this paper
has the higher average precision than that of TCHR and WCHR.
=
10
,
20
,
30
Table 1.
Comparison of average precision
1
p
2
p
3
p
Testing
image
TCHR
WCHR
WCHCR
TCHR
WCHR
WCHCR
TCHR
WCHR
WCHCR
Buses
68
70
74
60
62
65
50
56
58
Beach
59
61
65
54
55
56
48
50
55
Flowers
81
84.7
87.0
71.5
80.2
83.3
58.6
70.3
80.4
Horses
88.3
93.3
95.0
78.5
83.2
85.3
70.9
72.1
78.3
Food
73
76.1
81.3
67.5
73.8
75.2
60.2
62.9
67.1
A retrieval result of three retrieval methods is shown in Fig.1. The query image is
the upper-left corner image of each block of images, 30 images which are the most
similar to the query image have been output. The results are ranked in ascending order
of similarity to the query image from left to right and then from top to bottom. The
traditional color histogram can only reflect the statistical characteristics but its ability
to distinguish images is very limited, Thus, the false retrieval rate is very high, it can
be seen in Fig.1(a). Weighted color histogram retrieval (WCHR) considers signific-
ance of each pixel and the retrieval results are improved, such as Fig.1(b). While , the
output result in Fig.1(c) takes significance of each pixel and spacial color distributing
information of the same color into consideration, by this way, the retrieval results will
be much better in accordance with the visual perspective.
