Boosting Ensembles of Heavy Two-Layer Perceptrons for Increasing Classification Accuracy in Recognizing Shifted-Turned-Scaled Flat Images with Binary Features

A method of constructing boosting ensembles of heavy two-layer perceptrons is stated. The benchmark classification problem is recognition of shifted-turned-scaled flat images of a medium format with binary features. The boosting gain is suggested in two aspects. The earliest one is the ratio of minimal recognition error percentage among the ensemble perceptrons to the recognition error percentage performed by the ensemble. The second gain type is the ratio of minimal variance of perceptrons' recognition error percentages over 26 classes to variance of the ensemble's recognition error percentages over 26 classes. Both ratios increase as the number of perceptron classifiers in the ensemble increase. The ensemble of 36 classifiers performs with increased accuracy, where recognition error percentage is decreased for 33 %, and the variance is decreased for more than 50 %. Further increment of classifiers into ensemble cannot increase accuracy much as there is the saturation effect of the boosting gain. And the gain itself depends on the range of noise modeling object's distortions. Thus, the heavier perceptron classifier the less gain is expected.