CLASSIFICATION OF A GRANULAR PRODUCT USING HIGH-LEVEL FUSION OF VISION FEATURES

The characterization of a granular product based on image analysis can be a difficult problem because it often requires the combination of a large number of features of different natures extracted from the imag e. This problem of classification can be solved by two approaches. One of these approaches consists of aggregating the qualitative informati on which is obtained by considering each individual feature, as a virt ual sensor. This is a triple-step system: first, for each feature (i.e . virtual sensor), the samples are given a probability of belonging to a class (clustering); second, these probabilities are aggregated in o rder to give a global probability of the sample of belonging to each c lass (supervised neural network); third, the sample is assigned to the class which shows the maximal global probability. This procedure was applied to classify semolina samples. These were obtained by grinding wheat grains. Three classes were defined using three grinding roll gap s of 0.3, 0.4 and 0.5 mm, respectively. The average of correct classif ication was better than 80%. This methodology is particularly interest ing because it gives a very satisfactory result and is quite versatile new features added to the classification process require an update of one part of the procedure only. (C) 1997 Silsoe Research Institute.