Evaluating the Informativity of a Training Sample for Image Classification by Deep Learning Methods

A new approach to evaluating the informativity of a training sample when recognizing images obtained by means of remote sensing is proposed. It is shown that the informativity of a training sample can be represented by a set of characteristics, where each of them describes certain data properties. A dependence between the training sample characteristics and the accuracy of the classifier trained on the basis of this sample is established. The proposed approach is applied to various test training samples and their evaluation results are presented. When evaluating the training sample using the new approach, the process is shown to be much faster than that of training a neural network. This allows us to use the proposed approach for the preliminary estimation of a training sample in the problems of image recognition by deep learning methods.