Databases Reference
In-Depth Information
FIGURE 2.7: A comparison between subpattern and pattern subset distribu-
tion techniques.
partitioned into a number of modules known as windows. Each window is used
in a Hamming memory operation. The recognition results from each Hamming
memory are sent to a decision network to determine the final output from the
system. Mu et al. [41] extended the decoupled Hamming AM approach by
introducing a voting mechanism into the decision-making process.
In addition to the division of the input space into subspaces, pattern dis-
tribution techniques include a recognition process that is based on the atomic
pattern components that make up the entire pattern representation. For in-
stance, Khan and Mihailescu [2] proposed parallel pattern recognition in a
wireless sensor network (WSN) environment using a Graph Neuron (GN) ap-
proach. In their work, sensory data obtained from a sensor node was con-
sidered to be a component of the entire pattern represented by the network.
Subpattern distribution techniques allow recognition process to be performed
on minimal data, i.e., due to the size of the subpattern, the complexity is
low. Nevertheless, this technique is impossible to deploy in all deterministic
approaches. Some algorithms are highly cohesive, and the whole input space
must be included in its computations to obtain an optimum result.
2.4.2 Pattern Set Distribution
Pattern set distribution is a common approach in distributed pattern recog-
nition. It involves the distribution of separate input data sets to each of the
processing entities within the network.
Patterns are also distributed during the pre-processing stage of the clas-
sification/recognition process. For instance, Kokiopoulou and Frossard [42]
proposed a distributed support vector machine (SVM) approach for the clas-
sification of images within a sensor network. In this method, the input signal is
