Aphelion Neural Network Toolkit
A POWERFUL CLASSIFICATION SOFTWARE PRODUCT
BASED ON NEURAL NETWORK TECHNOLOGY
TO EFFICIENTLY CLASSIFY OBJECTS OF INTEREST
The AphelionTM Neural Network Toolkit (NNTK)
enhances AphelionTM Dev with an optional tool to
automatically classify objects of interest, based on a
supervised classification assigning objects to
categories or classes. It comes with Classifier
Builder, a front-end application to NNTK to easily and
seamlessly perform a manual classification to assign
objects of interest to known classes. Probability and
information based classifiers can be generated
automatically, freeing the user from the necessity of
specifying complex rules for object recognition and
classification.
This module includes tools to:
• Load ObjectSets in a classification project
• Define the different object classes
• Select the measurements to compute on all
loaded ObjectSets
• Select the attributes for use during the training
process
• Automatically compute the classifier parameters
The list of defined classes is saved in a project as an
XML file. The software takes as input one or more
Aphelion ObjectSets, and adds a new attribute
holding the class of the object after automatic
classification.
Classifier Settings
The second step of the classification process is the
settings of the classifier, i.e. the definition of user
parameters and the automatic settings used during
the training process.
This step includes the computation of the object
attributes selected by the user amidst all the
measurements available in Aphelion Dev such as
shape attributes (area, perimeter, Feret diameters,
etc.), texture attributes (Haralick parameters, etc.),
statistical attributes (pixel mean, minimum, maximum,
etc.). The list of selected attributes is saved in the
Classification project.
• Apply the classifier to any Aphelion ObjectSet
whatever its origin
• Perform a manual classification
• Run the classifier in Aphelion Dev
This module was developed in partnership with the
University of Caen (Normandy, France). It is based
on MONNA, a software product, originally conceived
by Dr. Olivier Lezoray.
Training Database Generation
The first step to be performed before defining the
classifier is to build a training database. Classifier
Builder provides tools to define the different classes,
to load the ObjectSets, to display the associated
objects, and to manually assign them a class.
All computed attributes are added to the loaded
ObjectSets, as well as the class attribute. The
ObjectSets can be saved and later imported into the
Dev version of Aphelion for further analysis.
By default, the Aphelion Neural Network Toolkit
proposes a classifier made of a set of single layer
neural network classifiers according to the MONNA
architecture. Combining multiple networks gives
better results than just using one multi-layer network.
Main benefits of Neural Network Tookit:
•
Easy to use graphical user interface for classification
•
Enhance Aphelion Dev with powerful set of tools for advanced Image Understanding
through the use of Aphelion Objectsets
• Optimized classification based on the neural network architecture
• Perform high-level classification in the fields of biology, cytology, quality control, optical
character recognition and any set of objects of interest when categories have to be
determined based on a large set of attributes.
A wizard is available to let the user quickly define the
parameters associated to the classifier: selection of
the attributes to process with, definition of the number
of neurons to be used to discriminate two classes,
and definition of the mode to select the attributes
used by the classifier. The number of parameters
defining the classifier was voluntary reduced to
simplify its definition. Other parameters, available by
program, are set such as the classifier is optimized
for most cases. The module is flexible enough to let
developers use as many layer and network as
wished, and define advanced parameters.
Comparison
techniques
with
other
classification
One of the main benefits of the Aphelion Neural
Network Toolkit is that it is more efficient than a
standard multi-layer neural network: less complex
architecture, faster training (partial training possible),
and better classification.
Table 1: Classification comparison between a multi-layer
neural network and Aphelion Neural Network Toolkit
Base
Liver Bupa
Pima
Ionosphere
Cancer
Wine
Vehicle
PageBlocks
Glass
Segmentation
OptDigits
MultiLayer
Network
71%
76,6%
90,1%
97,1%
97,3%
71,9%
85,4%
66,7%
82,4%
78,1%
Aphelion
NNTK
#Classes
71%
76,6%
90,1%
97,1%
100%
78,4%
90,1%
82,2%
87,8%
81,6%
2
2
2
2
3
4
5
7
7
9
Application fields
This module has been successfully used in the fields
of biology/cytology (i.e. cell analysis, etc.),
agriculture, quality control, optical character
recognition (i.e. license plate analysis, etc.), remote
sensing, etc. It is versatile enough to be applied to
other applications requiring advanced classification
tools.
The training step computes the weights of every
neuron to minimize the error on the classes assigned
by the classifier to the objects of the training
database. The values generated by the classifier
during the training step can be output in a text
window or displayed as a chart, to check the
convergence of the classifying process.
Finally, the classifier settings are saved in the project
to be easily retrieved for further classification.
Classification Process
Once the classifier is fully defined, it can be applied to
another ObjectSet loaded in the Aphelion Neural
Network Toolkit environment. The result of the
classification can be reviewed by browsing all the
object images and their assigned classes.
Supported Operating Systems
Windows XP, Windows Vista, and Windows 7 (32-bit
and 64-bit editions
Any Neural Network project can be loaded in
Aphelion Dev to apply the classifier to new
ObjectSets.
P. O. Box 6473
Monroe Twp., NJ 08831-6473
Tel: (877) 664-8772
Fax: (609) 944-8855
Web: www.AmerinexImaginng.com