M.Sc. Ospanov M.G.
A.Baitursynov Kostanay State University, Kostanay
Model image recognition.
The model is in the
article "Image recognition. Creating a model". Will test.
We create a new selection of 24 items. First 4re element are the same as in the
training set. The remaining options are different from the training set of
images:
Next load the data, and pass into
the procedure Recognize. It is averaged over each image is displayed in the
space of the principal components are weight w. Once known vector w is
necessary to determine which of the existing facilities it is the closest. To
do this, use the function dist (instead of the classical Euclidean distance in
pattern recognition problems is better to use another metric: Mahalanobis
distance). Is the minimum distance and the index of the object to which this
image is the closest.
On a sample of 24 objects shown
above 100% efficiency of the classifier. But there is one nuance. If we have to
apply to the input image, which is not in the source database, there will still
be calculated vector w and found the minimum distance. Therefore, a criterion
for O, if the minimum distance <O means the image belongs to the class of
recognizable, if the minimum distance > O, then such an image in the
database is not. The value of this criterion is chosen empirically. For this
model I chose O = 2.2.
Let's make a sample of persons who
are not in training and see how well the classifier will eliminate false
samples.
Of the 24 samples have four false positives. Ie efficiency was 83%.
Conclusion
In general, simple and original
algorithm. Once again proves that in spaces of higher dimension "hidden"
a lot of useful information that can be used in different ways. J combined with other advanced techniques eigenface can be applied to
improve the effectiveness of decision tasks.
For example, we have used as a
simple classifier distance classifier. However, we could use a better
classification algorithm, such as Support Vector Machine (SVM method) or a
neural network .
Literature:
1. onionesquereality.wordpress.com/