We present an image classification system based on a supervised learning method. The learning phase consists in an automatic rules construction: ≪ knowledge acquisition ≫ from training pixels is automatic. The obtained rules are classification ones : their conclusions are hypotheses about the membership in a given class. An inference engine uses these rules to classify new pixels. The building of the premises of the production rules is realized by linear correlation research among the training set elements. In this paper, we present and compare two methods of linear correlation searches : the first is done among all the training set without distinction of classes, and the second is an intra-classes search. An application to image processing in the medical field is presented and some experimental results obtained in the case of medical human thigh section are reported.