machine vision news

Machine Vision News

Vol. 8, 2003

Classification in Machine Vision

Separating objects into different classes is afundamental issue in all industry. This is particularly true in the field of machine vision. Recognition of faulty components, any grading or sorting of objects and analysis of multidimensional data all require decisions based on classification of objects.

An example of a classification task of separating stones from soil is presented on the figure below:

There are several issues complicating the classification task:

Necessary classification decisions are made based on information that may include dozens or even hundreds of ariables which may affect each other in a highly non-linear manner.
Relevance of the measured features is difficult and often impossible to be known beforehand.
In industrial real-time classification tasks, there is usually very little time for decision making.
Time and resources needed to build up a classification system are expensive.

Application Areas

Classification is needed in all areas of industry and business. Following list presents just a few of the possible application areas:

Machine Vision applications

Assembly line parts recognition
Grading of products
Automated medical diagnosis
Face recognition
Agricultural and geographical analysis
Chemical structure analysis

Speech and sound applications

Voice recognition
Sonar and radar signal detection
Seismographic recording

Text and document applications

Handwritten character recognition
Printed character recognition
Text and document classification

Benefits of Classification on Your Business

Several benefits can be achieved by using efficient classification in production. The emphasis of the benefits depends very much on the application area. Following is a list of the most typical benefits:

Improved yield in production
Reduced amount of negative customer feedback
Capability to handle situations where real-time demands are getting overwhelming for people
Finding new information about manufacturing processes

ClassifyIt - Sollution to Classification Problems

CIM VisionSolutions has developed a tool for solving any classification problem. ClassifyIt software contains all necessary components to create, teach and apply a classifier to your application. ClassifyIt supports all the latest technology classifiers: Including statistical and neural networks classifiers. That makes it easy to find the most effective classifier for each application.

ClassifyIt provides tools for easy creation of fast and accurate classifiers. All that is needed is a labelled archive of samples.

Teaching of classifiers is done in three steps:

Data preparation module prepares sample data for subsequent procedures.
Feature selection tools select powerful combinations of available features resulting in savings on needed feature calculation and improved classification performance.
State-of-the-art classification methods learn class boundaries straight from data without time taking tuning of classification rules resulting in ready classifiers quickly and with easy-to-use teaching user interface.

Phases of creation of classifiers is shown on figure below:

Data preparation

There are certain actions needed to prepare sample data for classification. Normalization techniques are needed to ensure optimal formation of feature space. Sample data must be properly divided to teaching-, testing- and validation sample sets to control the generalization capabilities of the classifiers. ClassifyIt provides all necessary algorithms needed for all-inclusive preparation of sample data. ClassifyIt data preparation dialog is shown on the following figure:

Feature Selection

On most occasions there are a lot of measurements made of the process for which the classification is implemented. Often many of these can be discarded while at the same time improving classification performance. A feature selection phase is therefore a prerequisite for efficient classification. ClassifyIt provides two top ranked methods for feature selection: SFFS(Sequential Forward Floating Selection) and GA(Genetic Algorithms) methods. An example of ongoing SFFS feature selection in ClassifyIt is presented on figure below:

Teaching and testing of Classifiers

After feature selection a classification method is chosen. The most suitable method for a specific application can never be known beforehand. Five state-of-the-art classifiers are therefore available for classification: kNN(k Nearest Neighbor), LVQ(Learning Vector Quantization), MLP(MultiLayerPerceptron) ,PNN(Probabilistic Neural Networks) and SOM(Self Organizing Maps) classifiers. An example of ongoing MLP classifier teaching process in ClassifyIt is presented on the following figure:

After a classifier has been taught it can be tested against validation sample set. Confusion matrix along with all relevant information about classification speed and accuracy is presented in a testing module.

An easy-to-use Windows GUI for classifier creation, wide selection of classifiers together with fast classification speed make ClassifyIt suitable for demanding real time industrial inspection environments. Runtime interface for classification is available in COM and DLL versions.

Contact Information:

CIM VisionSolutions Oy
Sammonkatu 6
FIN-90570 Oulu
Finland
www.cimvisionsolutions.com