Bias in Automated Image Colorization: Metrics and Error Types

Frank Stapel; Floris Weers; Doina Bucur

doi:10.48550/arXiv.2202.08143

Bias in Automated Image Colorization: Metrics and Error Types

Frank Stapel, Floris Weers, Doina Bucur

Research output: Working paper › Preprint › Academic

Abstract

We measure the color shifts present in colorized images from the ADE20K dataset, when colorized by the automatic GAN-based DeOldify model. We introduce fine-grained local and regional bias measurements between the original and the colorized images, and observe many colorization effects. We confirm a general desaturation effect, and also provide novel observations: a shift towards the training average, a pervasive blue shift, different color shifts among image categories, and a manual categorization of colorization errors in three classes.

Original language	English
DOIs	https://doi.org/10.48550/arXiv.2202.08143
Publication status	Published - 16 Feb 2022

Keywords

cs.CV
68T45
I.4.4

Access to Document

10.48550/arXiv.2202.08143

2202.08143v1Submitted version, 3 MB

Cite this

@techreport{b5f509a4ee6b4545b83da28a66b2d3c5,

title = "Bias in Automated Image Colorization: Metrics and Error Types",

abstract = "We measure the color shifts present in colorized images from the ADE20K dataset, when colorized by the automatic GAN-based DeOldify model. We introduce fine-grained local and regional bias measurements between the original and the colorized images, and observe many colorization effects. We confirm a general desaturation effect, and also provide novel observations: a shift towards the training average, a pervasive blue shift, different color shifts among image categories, and a manual categorization of colorization errors in three classes.",

keywords = "cs.CV, 68T45, I.4.4",

author = "Frank Stapel and Floris Weers and Doina Bucur",

note = "5 pages, 8 figures",

year = "2022",

month = feb,

day = "16",

doi = "10.48550/arXiv.2202.08143",

language = "English",

type = "WorkingPaper",

}

TY - UNPB

T1 - Bias in Automated Image Colorization

T2 - Metrics and Error Types

AU - Stapel, Frank

AU - Weers, Floris

AU - Bucur, Doina

N1 - 5 pages, 8 figures

PY - 2022/2/16

Y1 - 2022/2/16

N2 - We measure the color shifts present in colorized images from the ADE20K dataset, when colorized by the automatic GAN-based DeOldify model. We introduce fine-grained local and regional bias measurements between the original and the colorized images, and observe many colorization effects. We confirm a general desaturation effect, and also provide novel observations: a shift towards the training average, a pervasive blue shift, different color shifts among image categories, and a manual categorization of colorization errors in three classes.

AB - We measure the color shifts present in colorized images from the ADE20K dataset, when colorized by the automatic GAN-based DeOldify model. We introduce fine-grained local and regional bias measurements between the original and the colorized images, and observe many colorization effects. We confirm a general desaturation effect, and also provide novel observations: a shift towards the training average, a pervasive blue shift, different color shifts among image categories, and a manual categorization of colorization errors in three classes.

KW - cs.CV

KW - 68T45

KW - I.4.4

U2 - 10.48550/arXiv.2202.08143

DO - 10.48550/arXiv.2202.08143

M3 - Preprint

BT - Bias in Automated Image Colorization

ER -

Bias in Automated Image Colorization: Metrics and Error Types

Abstract

Keywords

Access to Document

Fingerprint

Cite this