METHOD FOR 3D RECONSTRUCTION OF A HUMAN HEAD TO OBTAIN A RENDER IMAGE OF A PERSON Russian patent published in 2022 - IPC G06T15/00 G06T19/00 

FIELD: image3D reconstruction.

SUBSTANCE: invention relates to a method for 3D reconstruction of a human head to obtain a render image of a person. The method uses a single source image, wherein the face shape is extracted from a single source image, head orientation, facial expression are extracted from an arbitrary target image, and the method is carried out on a device having a CPU, an internal memory with images, RAM, and the method comprises steps in which : a) reading a single original image from the device memory and encoding, by means of the first convolutional neural network, the original image into a neural texture of the same spatial size as the original image, but with an increased number of channels, which contains local personal details; b) parallelly estimating face shape, facial expression and head orientation by the retrained DECA system with respect to source and target images, and provide an initial mesh as a plurality of faces and a plurality of initial vertices; c) providing a predicted mesh for the non-facial parts of the head mesh, and at the provisioning stage: render the initial mesh into an xyz-coordinate texture, chain the xyz-coordinate texture and the neural texture, process using the second network into a latent geometric map, bilinearly filter the latent geometric map with using texture coordinates to obtain a vertex-dependent feature, decoding the vertex-dependent feature by a multilayer perceptron to predict a 3D displacement for each vertex, adding the predicted 3D displacement to the initial vertices to obtain a 3D reconstruction of a human head as a predicted mesh; d) rasterizing the 3D reconstruction of the human head to process the result through a third network to obtain a render of the human image.

EFFECT: invention provides the possibility of generating a realistic image of the human head.

1 cl, 3 dwg, 1 tbl