IMAGE PROCESSING SYSTEM AND METHOD USING DEEP NEURAL NETWORKS Russian patent published in 2021 - IPC G06N3/02 G06N3/08 G06T5/20 

FIELD: physics.

SUBSTANCE: invention relates to computer engineering for image processing. System and method implement in-depth training on a mobile device to create a convolutional neural network (CNN) for real-time video processing, for example, for hair dyeing. Images are processed using CNN to determine corresponding pixel mask of hair. Corresponding object masks can be used to determine which pixel needs to be changed when correcting pixel parameters, for example, to change colour, lighting, texture, and so forth. CNN may include a (pre-trained) network for classifying an image, adapted to obtain segmentation isolation. CNN can be trained for segmenting an image (for example, using coarse segmentation data) in order to minimize loss of compatibility of extraction and image gradients. CNN may additionally use skip-connections between corresponding layers of coder stage and decoder stage.

EFFECT: technical result is higher resolution of output image data.

82 cl, 6 tbl, 10 dwg