Real-world evaluation of artificial intelligence-based color corrections for social media content creators

Abstract

Brands strive to maintain consistent brand color representation across many digital channels, including social media platforms. This is a challenging goal given the varying real-world circumstances in which brand imagery is produced and shared. The ColorNet neural network tool was developed to automatically target and correct brand colors in imagery without altering non-brand colors. Previously, it was successfully applied to live sports broadcast footage. An open question is whether ColorNet can improve the accuracy, as measured by ∆E00, of brand color representa- tions in still photographs and photographs taken from videos and gifs posted to social media platforms such as Instagram and Twitter. To test this question, we collected a set of posts containing imagery from social media created by Clemson Athletics’ social media accounts. We corrected the representation of Clemson orange in these images using ColorNet. After selecting pixel values corresponding to brand colors in each media piece, we demonstrate that ColorNet improves Clemson brand color accuracy across social media channels and for various media characteris- tics. Despite our observation that brand color representation varies significantly across media types and lighting conditions, the improvement in color representation from ColorNet was relatively consistent. We also showed that ColorNet has a comparatively minor impact on the color representation of skin tones.