Bias in Automated Image Colorization: Metrics and Error Types

TL;DR

This paper investigates biases in automated image colorization, introducing new metrics for local and regional bias, and categorizing common error types to better understand and evaluate colorization quality.

Contribution

It presents novel bias measurement techniques and error categorization methods for assessing GAN-based image colorization models.

Findings

Desaturation effect observed in colorized images

Blue shift is a common bias across results

Color shifts vary among different image categories

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.