US 11,880,961 B2
Spectral hint-based inpainting of images
Xingqian Xu, Kirkland, WA (US); Shant Navasardyan, Yerevan (AM); Vahram Tadevosyan, Yerevan (AM); Andranik Sargsyan, Yerevan (AM); and Humphrey Shi, Princeton Junction, NJ (US)
Assigned to PICSART, INC., Hallandale Beach, FL (US)
Filed by Picsart, Inc., Hallandale Beach, FL (US)
Filed on May 12, 2023, as Appl. No. 18/196,640.
Claims priority of provisional application 63/341,799, filed on May 13, 2022.
Prior Publication US 2023/0368341 A1, Nov. 16, 2023
Int. Cl. G06K 9/00 (2022.01); G06T 5/00 (2006.01); G06T 5/10 (2006.01); G06T 5/20 (2006.01)
CPC G06T 5/005 (2013.01) [G06T 5/10 (2013.01); G06T 5/20 (2013.01); G06T 2207/20048 (2013.01); G06T 2207/20081 (2013.01)] 20 Claims
OG exemplary drawing
1. A computer-implemented method comprising:
obtaining an input image tensor of image data, the input image tensor comprising a known region and a missing region, wherein the known region comprises a plurality of known pixel data units and the missing region comprises a plurality of unknown pixel data units;
transforming the input image tensor of the image data into a spectral tensor of the image data wherein one or more of: row values or column values, of each matrix in the input image tensor are provided as a sequence of pixel data values in a spatial domain for which corresponding frequency values are determined in a frequency domain;
modifying the spectral tensor of the image data to generate a filtered spectral tensor of the image data at least by applying a filter, having a plurality of weights, on the spectral tensor;
wherein one or more weights of the plurality of weights of the filter are determined at least in part by a machine learning algorithm;
generating output image data, at least in part, by transforming the filtered spectral tensor into the spatial domain;
using the known region of the input image tensor and based, at least in part, on a region, of the output image data, that corresponds to the missing region of the input image tensor, generating a final output image.