Up: Samplex Table of Contents Previous: Statistical Pattern Recognition Next: Classifying the image

Neural Network Classifier

The BLC2 neural network classifier was developed for color classification. During the learning phase corrections are made to the classifier for each iterations using a correction coefficient. The number of iterations, the learning coefficient and the error objective can be specified in the control window. Default values are 20 iterations, a learning coeficient of 1.5, and a learning objective of 0.1% error.

Figure 10: nSPR control window

BLC2 is using a recursive dichotomy of the color space. Thus the order of the dichotomy has to be set-up prior to learning. A set-up window allows the user to specify the order by clicking on the classes button in the right order. Classes' statistics are displayed in the window to help making a decision.

Figure 11: Class order window

Usually a good bet is to order classes by intensity. In this example class 0 is the darkest. If the network does not converge, the order of the classes can be changed. To set up the right combination, some 3D graphs are needed.

Once the order is set, the learning begins. After each iterations some information (R error, error distribution) are displayed in the control window.

Errors are also displayed after each iterations in the original image.

Figure 12: Interactive display of errors

This display method helps detecting what classes are at fault, and what sampled area generated errors. It allows the user to correct its sampling.

After the number of iterations or the learning objective is met, the confusion matrix is displayed.

Figure 13: Learning results

Up: Samplex Table of Contents Previous: Statistical Pattern Recognition Next: Classifying the image