METHOD OF COMPRESSION OF IMAGE VECTOR Russian patent published in 2018 - IPC G06T1/40 G06T9/00 G06N3/08 

FIELD: computer engineering.

SUBSTANCE: invention relates to computer engineering. Method for training an auto-encoder for compressing an image vector includes creating a reference vector of target image characteristics based on an image vector, wherein the reference vector includes information about the target image characteristics from the image vector; compressing the image vector using an auto-encoder to obtain a compressed image vector based on the image vector; decompressing the image vector using an auto-encoder to obtain a lossy image vector based on the image vector; creation of a vector of target characteristics of the image with losses based on the loss vector; comparison of a reference vector of target image characteristics with a vector of target image characteristics with losses by determining the discrepancy parameter and using the discrepancy parameter to train the autocoder so that the loss of information in the image vector with loss, associated with the target characteristics, are reduced due to the increased loss of information associated with the additional characteristics of the image.

EFFECT: technical result is to provide selective minimization of loss of certain image characteristics.

21 cl, 5 dwg