Workflow from Scientific Research

Overview of our proposed text appearance adaptation network. It is composed of a geometry transformation module (GTM) (left) and a color harmonization module (CHM) (right). Given an input triplet consisting of a patch plain-text image $P_{pt}$, a reference rectangle $Rect$, and a background image $I_{bg}$, the GTM learns to place the text with a realistic perspective. Then, the CHM takes the composite image $I_{comp}$ and text mask $I_{ttA}$ as inputs, and outputs a synthetic text image $I_{out}$ with a harmonious color.

DOI