Angiogram 1 original

1
Applications of Cellular Neural Networks to Image
Understanding Experimental results on
Shape/motion Estimation and Recognition
2
Experimental Results
Nerwork
Figure Estimating shape and motion using
Cellular Neural
(
input image on the left, result on the right)
3
SELECTING FEATURE VECTORS FOR RECOGNITION
image
sequence
optical
flow
features of
flow
field
rearranged
features
of time series
4
Experiments using CNN as a associate memory
For experiments we are currently using images of
the Columbia
image database. We take image sequences of 36
images each of
several selected objects. To speed up the flow
computation and to
handle the data amount, we reduced the image
resolution to 32x32
pixels.
Figure Some images of the Columbia image database
normalized
gray images.
For a better visualisation they are shown as
We show the different features in x-direction,
the time in y-direction.
The associative memory is used to restore
incomplete sequences and
to classify them.
Figure Feature vectors of five objects
We find that using 8 features out of the full set
of 13 features leads to
excellent discrimination.
Further experiments will be done concerning
different resolutions of
the images and concerning variable lengths of the
image sequences.