Once classification has been applied to all pixels of the image, we wish to estimate the quality of the classification obtained.

This estimation consists in determining the percentage of correctly classified pixels, i.e. assigned to the class c to which they really belong. But the only knowledge available on the true class of membership of the pixels is reduced to the samples having been used to learn to recognize each class. We know that all the pixels composing the samples of a class c belong to this class. To estimate how many pixels are correctly classified, it is enough to compute how many pixels of the samples of class c were classified in class c.

The rate of recognition of class c is the ratio between:

• the number of pixels correctly classified in class c and
• the number of pixels constituting the samples of the class c .

But learning to recognize a class from examples, and testing classification with the same examples, gives an overestimation of the quality of the classification. In effect, the classifier is obviously most efficient in the area of the samples.