Information-extremе machine learning for object identification on the terrain

The article deals with the usage of a SURF local descriptor of key fragments to create a global descriptor BoF for objects of interest on terrain within task of recognition of armored technique in the controlled territory using images of air reconnaissance. The method of optimization of the input mathematical description of the information-extreme classifier trained on dataset, which consist of global descriptors BoF, is proposed. The optimal in information understanding dimensionality of a global descriptor is determined by iterative procedure, which includes k-means clustering of key fragment's SURF-vectors, training dataset creation and information-extreme machine learning of the classifier. The offered algorithm of information-extreme machine learning implements the adaptive coding of values of primary features using multilevel system of control permits, and creation of hyperspherical containers of classes in binary space of secondary features with sequential optimization procedures. It was suggested to use the rated modification of S. Kulbak's information measure, which is a function of false omission rate and positive predictive value of decision-making and it also allows machine learning on imbalanced dataset.