CLASSIFICATION OF FRUITS BY A BOLTZMANN PERCEPTRON NEURAL NETWORK

Classification of fruits by machine vision is problematic in two respects: (a) Most of the sorting criteria are "fuzzy", because the class membership can not be quantified precisely. The reference classification is subjectively determined by a trained panel of inspectors, that often disagree as to the class of the fruit. (b) The statistics of the classification criteria vary with harvest time and from orchard to orchard, so the classifier must be easy to re-train. Using digital color imaging hardware and a BPN based classifier we developed a system for sorting fruits that can address these problems. It naturally accepts "fuzzy" or "soft" labeling at train-time and can be tuned to provide various levels "soft" or "fuzzy" decisions at run-time. The power of the BPN was demonstrated by a synthetic dataset, indicating that the BPN can create intricate non-linear discriminant functions, even when the classes are noncontinguous and the training set is relatively small. Simulated sorting experiments of apples, into "red" and "green" categories showed that the system can emulate the decisions of a panel of human sorters quite well.