It was compos ed of three steps.
1. Convol ution layer:
The goal is to extract feature point. For
that, it uses many filters which are 3x3 or
5x5 matrix usuall y. As it uses various filter s,
it genera tes many convol ution images.
And ydo you know how applyi ng convol
ution kernel s(filter) er) to the origin al image?
It is simple. For exampl e,
1 2 1 3
2 3 4 3
4 3 4 5
The upper is image binary. Let's assume that
each digit repres ents gray color value. And the
matrix below repres ents convol ution kernel.
3 2 1
2 3 1
2 3 2
In the origin al image binary, (2,2) has
value 3. And (3,2) has value 4.
Let's calcul ate the value of (2,2). You
can crop the neighb or values of (2,2). It
is like below.
1 2 1
2 3 4
4 3 4
Let's multip ly each positi on value of
the upper cropped matrix by each positi on value
of the convol ution kernel. And add each multip lied
1x3 + 2×2 + 1×1 +
2×2 + 3×3 + 4×1 +
4×2 + 3×3 + 4×2
So the value 50 is the (1,1)'s convol
As you has already known, the first row
and column can't be applie d. So, the genera
ted convol ution matrix is a little smaller size than
the origin al o ne. e. Two or five pixels are
reduce d. Usually the border area in the image is not import
ant, so it is ignore d.
2. Subsam pling.
Then, it reduced convol ution image size. It
may be for operat ion effici ency with my
guessi ng. As the key featur es and their geomet
ry relati on are kept, subsam pling is safe.
3. Iterate the upper 1 and 2 steps.
4. You can get very small size convol
ution matrixs and their number would be numero us.
With these numero us matrix values, you can
apply classi fication on algori thm.
And the result will be recogn ized object
If you are traini ng with given data,
you are making convol ution network model. And if
you are predic ting with given image, you are applyi
ng the given image to the convol ution network model.
원본 : http://blog.daum.net/jhmoon77/17456608