Recognition Methodology
During the training session, the network is presented with patterns that
are recorded by the neurons at the first layer. They are retained by parts
with each neuron recording a 4×4 block. The
same block is retained by 25 shift neurons, each recording a specific shift.
During recognition phase, the first-layer neurons try to match the subpatterns
they have retained, If any of the 25 shift neurons fires, the second-layer
neuron is also feeding fires, announcing the detection of the subpattern.
All of these 16 subpattern neurons work together to detect the complete
pattern. If all 16 subpattern neurons fire, they announce the detection
of a pattern. However, it is possible to define a reliable range such as
12 or 13 around this pattern neuron. This neuron‘s output goes to the output
layer that provides the classification.
The training patterns of the water break defects are shown in
Fig. 4. Results obtained from experiments show that the networks could
identify most of the water break defects on the copper plate since the
true defect patterns take geometric alterations from our training patterns.
We also create other training patterns to detect the defects of spotting
out, scratching, pitting and smut. From the type
of the defects: Sphere, Cube, Fiber, and Flake, we can also
realize the particle source and control the contamination.
Fig. 4 the training patterns
nextindexback