A Zero-Shot Image Enhancement Method Based on Human Skin Tone Reference

• Diulhio Oliveira Venturus - Innovation & Technology
• Alexandre Lopes Venturus - Innovation & Technology / UNICAMP
• Vitor Galassi Luquezi Venturus - Innovation & Technology
• Carlos Eduardo Nogueira Venturus - Innovation & Technology
• Mauricio Schiezaro Venturus - Innovation & Technology
• Dennis Liu Compal - Electronics Inc.
• Helio Pedrini UNICAMP

Resumo

In this work, we introduce a zero-shot method for color enhancement in images containing people. This method is based on photo edition techniques used by professional editors, where these professionals adjust the RGB curves, which affects the image colors globally. The hypothesis adopted in this investigation is that, when there are people present in an image, they become the most important element of the scene, so one of the main aspects of color adjustment is to make their skin tones look more realistic and natural, which should please most of the viewers. The proposed method employs a pipeline of classical algorithms of computer vision to mimic the process applied by professional photo editors. Experiments performed on the Adobe-MIT 5k dataset, a widely used dataset on image enhancement, show that our approach has similar performance compared to supervised data-driven state-of-art methods.

Palavras-chave: Image enhancement, skin tone, deep learning, zero-shot method