Imbalance Satellite Image Colorization with Semantic Salience Priors

I. Overview

Satellite-image colorization is under-constrained: many colors may be plausible for a grayscale input, while dominant land-cover types can bias a model toward frequent outputs. This work introduces semantic salience priors to emphasize meaningful scene regions during color prediction.

II. Key Contributions

• Incorporates semantic salience into satellite-image colorization.
• Addresses imbalance among common and infrequent scene content.
• Guides color prediction with higher-level scene information in addition to local intensity.

III. Methodology

The supplied overview shows a learned colorization pipeline augmented by semantic and saliency cues. These priors reweight or refine visual features so the decoder can assign colors according to scene content rather than frequency alone.

IV. Research Focus

The project explores how semantic structure can stabilize colorization when training data contain uneven land-cover and color distributions.

@inproceedings{zheng2021imbalance,
  title={Imbalance satellite image colorization with semantic salience priors},
  author={Zheng, Chengyu and Fu, Yu and Zhao, Zian and Wang, Chenglong and Nie, Jie},
  booktitle={Twelfth International Conference on Graphics and Image Processing (ICGIP 2020)},
  volume={11720},
  pages={546--552},
  year={2021},
  organization={SPIE}
}