METHOD OF TRAINING A CONVOLUTIONAL NEURAL NETWORK FOR IMAGE RECONSTRUCTION AND A SYSTEM FOR FORMING AN IMAGE DEPTH MAP (VERSIONS) Russian patent published in 2019 - IPC G06T7/593 H04N13/10 

FIELD: image processing means.

SUBSTANCE: group of inventions relates to the field of image processing. Method comprises the steps of: obtaining a stereoscopic pair image including the original left and right images; each of the images is transmitted to the input of the corresponding Siamese convolutional neural subnetwork to estimate the depth map; processing the supplied images to obtain high-level features maps and back depth maps for the left and right images; obtained maps of high-level features are transmitted to input of convolutional neural subnetwork for evaluation of chamber parameters; processing obtained maps of high-level features to obtain camera parameters, by means of which original left and right images were taken; applying affine transformation to reverse depth maps for the left and right images, respectively, taking into account the obtained camera parameters to obtain disparity maps for the left and right images; bilinear interpolation samples are used to obtain a reconstructed right and left image, recovered disparity maps for a right and left image, an auxiliary right and a left image, generating a loss function for reconstructed images based on original left and right images, reconstructed left and right images, disparity maps, recovered disparity maps for left and right images and auxiliary images.

EFFECT: increased accuracy.

12 cl, 8 dwg