Databases Reference
In-Depth Information
FIGURE 8.7: Edge map generated by applying Sobel's edge detection tech-
nique to an original gray-scale facial image.
with an n-dimensional feature vector,{p
1
, p
2
, . . . , p
n
}, where p
i
represents the
normalized percentage of color pixels that corresponds to each color element in
an image. Nascimento and Chitkara [93] proposed an alternative approach for
color distribution representation using a global binary signature scheme. It is a
compact form of the existing GCH that uses binary bit-strings as a signature.
This signature is an abstract representation of the image's color distribution.
The bit-strings are of a pre-determined size, which makes it ideal for use in
DHGN binary pattern representations.
8.2.1.2
Edge Detection for Structural Information
Edges provide important spatio-structural information for image recogni-
tion. This multi-feature DPR scheme includes edge detection in the color-
based recognition process. The outputs from the edge detection process are
represented as an edge map. Figure 8.7 shows the transformation of a gray-
scale image to the corresponding edge map using Sobel's edge detection tech-
nique.
With the ability to capture and convert the two main features of an im-
age, i.e., colors and edges, into binary patterns, the distributed multi-feature
recognition scheme is able to apply a highly scalable single-cycle learning tech-
nique for binary patterns in a computational network for multi-feature pattern
recognition. Any number of features can be included in the scheme provided
a separate network is available for each feature (see Figure 7.2).
An interesting characteristic of the DHGN implementation for multi-feature
recognition is the constant recall time for each feature. This characteristic is
independent of the number of input patterns presented. Furthermore, the
scheme minimizes the recall time. Figure 8.8 shows the overall store/recall
times for each DHGN subnet in a face recognition simulation using 1000 im-
ages. In a simulated computation network, each DHGN subnet processed all
