US 12,423,948 B2
Systems and methods for predicting long-tail colors in images
Qiuyu Chen, San Jose, CA (US); Quan Hung Tran, San Jose, CA (US); Kushal Kafle, Sunnyvale, CA (US); Trung Huu Bui, San Jose, CA (US); Franck Dernoncourt, Seattle, WA (US); and Walter W. Chang, San Jose, CA (US)
Assigned to ADOBE INC., San Jose, IA (US)
Filed by ADOBE INC., San Jose, CA (US)
Filed on Jul. 26, 2022, as Appl. No. 17/814,921.
Prior Publication US 2024/0037906 A1, Feb. 1, 2024
Int. Cl. G06V 10/764 (2022.01); G06V 10/56 (2022.01); G06V 10/774 (2022.01)
CPC G06V 10/764 (2022.01) [G06V 10/56 (2022.01); G06V 10/774 (2022.01); G06V 2201/10 (2022.01)] 17 Claims
OG exemplary drawing
 
1. A method for color prediction, comprising:
receiving an image that includes an object comprising a color;
generating a color vector based on the image using a color classification network, wherein the color vector comprises a color value corresponding to each of a plurality of colors, and wherein the color classification network performs a convolution operation on the image to obtain the color vector;
generating a bias vector based on identifying a debiasing factor, comparing the color vector to each of a plurality of center vectors, and computing a distance function between the color vector and each of the plurality of center vectors based on the debiasing factor and the comparison, wherein each of the plurality of center vectors is generated by computing an average vector of a plurality of color feature vectors corresponding to a color of the plurality of colors; and
generating an unbiased color label for the image based on the color vector and the bias vector, wherein the unbiased color label accurately indicates the color of the object based on Classifier Sparse Encoding (CSE) for the plurality of colors by removing classifier bias and feature bias from the color vector with the bias vector.