COMPUTATIONALLY EFFICIENT MULTI-CLASS IMAGE RECOGNITION USING SUCCESSIVE ANALYSIS OF NEURAL NETWORK FEATURES Russian patent published in 2019 - IPC G06K9/62 G06N20/00 

FIELD: computer engineering.

SUBSTANCE: invention relates to the computer engineering. Image recognition method, in which a deep convolutional neural network (CNN) vector is obtained by a deep image vector of an input image; method includes applying a PCA transformation (principal component method) to a given vector to obtain a sequence of main components of the input image; dividing sequence of main components into a predetermined number of adjacent parts, each of which relates to a different granularity level; and attaching part of sequence related to this granularity level to initial empty sub-sequence of main components of input image, calculating distances between subsequence and corresponding subsequences of main components of standards from plurality of candidate solutions, estimating ratios of minimum distance to other distances, eliminating standards with ratios below threshold from plurality of candidate solutions, and if set of candidate solutions includes only one class in references, input image is identified as belonging to this class.

EFFECT: high efficiency of recognizing images.

19 cl, 11 dwg, 3 tbl