白丝美女被狂躁免费视频网站,500av导航大全精品,yw.193.cnc爆乳尤物未满,97se亚洲综合色区,аⅴ天堂中文在线网官网

Method and system for color representation generation

專利號(hào)
US11176715B2
公開(kāi)日期
2021-11-16
申請(qǐng)人
THE GOVERNING COUNCIL OF THE UNIVERSITY OF TORONTO（CA Toronto）
發(fā)明人
Maria Shugrina; Amlan Kar; Sanja Fidler; Karan Singh
IPC分類
G06T11/00; G06T15/50; G06T11/40
技術(shù)領(lǐng)域
color,sail,colors,in,image,neural,sails,alpha,patchwork,masks
地域： Toronto

摘要

There is provided a system and method for color representation generation. In an aspect, the method includes: receiving three base colors; receiving a patchwork parameter; and generating a color representation having each of the three base colors at a vertex of a triangular face, the triangular face having a color distribution therein, the color distribution discretized into discrete portions, the amount of discretization based on the patchwork parameter, each discrete portion having an interpolated color determined to be a combination of the base colors at respective coordinates of such discrete portion. In further aspects, one or more color representations are generated based on an input image and can be used to modify colors of a reconstructed image.

說(shuō)明書

Initially, the present inventors tried palette learning on random 32×32 patches extracted from images rescaled to 512×512 in the combined target and dirty training sets but determined that random patches mostly contain muted solid colors from background and low-frequency areas of the image; making it not only hard to train, but to evaluate performance on such data. To overcome this, it was observed that histogram entropy computed over the patch colors is a rough indicator of color complexity present in the patch (FIG. 12A). The first row of FIG. 12A shows sample entropy and the second row shows colorfulness of image patches and images. By examining patches of various entropy levels, “easy” patches with entropy <1.5 were designated as easy, patches with entropy over 3 as “hard”, and the rest as “medium”. The tests were split accordingly to better evaluate performance. See FIGS. 12A and 12B for patch entropy distributions in the datasets. As shown in FIG. 12B, colorfulness can be used to weigh unlabeled datasets to better approximate target data distribution, and entropy to designate various levels of difficulty for palette learning test sets for the first neural network.

In order to encourage more difficult patches to occur in the training set during target patch selection, the random patch selector was center biased, as the central area of the image is more likely to contain the more detailed main subject. In addition, patches of random size were generated and then the distribution was rescaled. This generally makes the training set of the first neural network more analogous to the region-based input seen during training of the second neural network.

權(quán)利要求

1
微信群二維碼
意見(jiàn)反饋