Digital Signal Processing Reference
In-Depth Information
Image Type
Sample Images
Original
input image
Some of the
extracted
image com-
ponents
Fig. 4.
Random Location Subwindows (RLS) placements displayed on the input images (Upper
row) and some of the image components extracted based on the RLS (Lower row)
3.3
Visual Content Matching (CM Layer)
This layer is to perform the calculation of feature vector closeness. The closeness
between the feature vectors generated from each elementary image with the feature
vectors obtained from each concept registered in OF KB will be calculated. We pro-
pose Cosine Similarity to measure the closeness between these two feature vectors.
Assumed the type
k
th
of feature vector extracted from component
e
i
is denoted as
, which is one of the elements obtained from
Ψ
. Also, assumed type
k
th
and
j
th
feature vector belonged to a concept
c
; obtained from OF KB, is denoted as
.
,
The Cosine Similarity of these two vectors is defined as
where
,
.
||
||
(3)
||||
Consequently, for each of the image components will draw a similarity value of
S
,
with respect to each of the training items that belonged to a specific registered con-
cept. Let
Ψ
CM
be the outcome of the matching operation from this layer, we have:
Ψ
,
1,
;1,Ψ
;1,
(4)
where
known as the total concepts available in OF KB as shown in eq. (1).
3.4
Objects Recognition (OR Layer)
The operations from previous layers have enabled a similarity value being drawn to
each registered concepts in OF KB. Further analysis based on these values is required
to draw a decision for object identification. We proposed a method called Multi-Level
Thresholds (MLT) to analyse the similarity values and derive a conclusion. The MLT
model defined as below:
