METHOD FOR REGULARIZING GRADIENTS WHEN TRAINING DEEP NEURAL NETWORK IN ALGORITHM FOR INCREASING FRAME RATE OF VIDEO IMAGE WITH IMAGE DEFORMATION OPERATION AND DEVICE AND MACHINE-READABLE MEDIUM REALIZING SAID METHOD Russian patent published in 2025 - IPC H04N7/14 G06T3/4007 G06V10/40 

FIELD: processing of video images.

SUBSTANCE: invention relates to video processing, and in particular to time interpolation of video frames to increase frame rate (FPS) of video. Disclosed is a video frame rate conversion (FRC) method comprising steps of: obtaining at least two video key frames (KF) from an input video frame sequence; generating at least one interpolated frame for adding to the video image frame sequence based on at least one of the KF, wherein the formation of at least one interpolated frame comprises steps of estimating motion vectors (MV) for at least one pixel of each block of the interpolated frame using a trained motion estimation algorithm (ME), trained using loss function gradients regularization by selection of estimated motion vectors from optical flow and modification of fractional parts of coordinates of selected motion vectors with back propagation of error without taking into account of said modification and/or using finite-difference approximation of spatial derivatives of key frames (KF) brightness with back propagation of error; obtaining at least one interpolated frame by performing motion compensation using said at least one key frame (KF) and motion vectors (MV) obtained by said estimation of motion vectors; adding the obtained at least one interpolated frame to the sequence of frames of the video image; and outputting a sequence of video frames with an increased number of frames. Invention also discloses a device and a machine-readable medium which realizes said method.

EFFECT: high quality of converting the frame rate of a video image.

1 cl, 9 dwg, 1 tbl