Database Reference
In-Depth Information
H
1
t
H
2
H
3
H
4
H
5
f
MHI
=
(
l
)
,
(
l
)
,
(
l
)
,
(
l
)
,
(
l
)
(3.21)
where
H
j
(
l
)
is the histogram generated from the coding label of the
j
-th subband,
l
=
1
,
2
,...,
L
, and the sub-codebook size
L
=
256. In Fig.
3.6
a, let the function
m
×
m
Q
L
be the function that quantized the value of the
i
-th input
vector of size
m
-by-
m
to its bin. The histogram model of each subband consists then
of the
L
values of the
L
bins of the histogram
H
(
V
i
)
:
R
ₒ
1
,...,
t
.
(
l
)=[
H
(
1
)
,...,
H
(
l
)
,...,
H
(
L
)]
The value of the
l
-th bin is calculated by:
N
V
i
=
1
ʴ
(
Q
(
V
i
)
−
l
)
,
H
(
l
)=
l
∈ {
1
,...,
L
}
(3.22)
where
is the Kronecker delta function, and
N
V
is the total number of vectors in
the subband.
ʴ
3.3.3
Experimental Result
The pseudo RF for adaptive image retrieval outlined in Fig.
3.3
was applied for the
retrieval of compressed images. Feature space
1
is characterized by the compressed
domain features (extracted from the DCT or WT compressed images). For relevance
classification in feature space
F
2
, a Gabor wavelet transform technique [
91
,
92
]was
employed. The experiments were designed to compare the performances of four
methods: non-adaptive CBIR, user-controlled RF retrieval, pseudo RF retrieval, and
semiautomatic retrieval. The experimental results were obtained from two image
databases: DB1, the Brodatz database, which contains 1,856 texture images; and
DB2, distributed by Media Graphic Inc. [
93
], consisting of nearly 4,700 JPEG color
images covering a wide range of real-life photos, with a completely open domain.
These are typical, medium-size databases that are potentially remotely accessible
through Internet environments, without advanced indexing of the stored images
[
76
].
For the DB1 test set, the visual descriptor used both WM [cf. Eqs. (
3.19
) and
(
3.20
)] and MHI feature [cf. Eq. (
3.21
)] representations, for the characterization
of the wavelet-compressed images. The WM descriptor was obtained before the
invert-DWT process of the wavelet-baseline coder, whereas the MHI descriptor was
obtained before VQ-decoding of the WT/VQ decoder. For the DB2 test set, the
visual descriptors used are the energy histograms which were extracted directly from
the compressed JPEG images after the entropy decoding process.
F
Search WWH ::
Custom Search