Adaptive Fine-Grained Region Matching for Image Harmonization.

Liuxue Ju,Chengdao Pu,Fang Gao,Jun Yu
DOI: https://doi.org/10.1007/978-3-031-46311-2_1
2023-01-01
Abstract:Image harmonization aims to generate composite images that are visually consistent by adjusting the foreground to be compatible with the background. However, previous image harmonization methods overlook the fact that in a real image, the appearance (e.g., illumination, color temperature, saturation, hue, and texture) of different regions can vary significantly depending on content and position. For each foreground region, the background regions related to it should be taken as major references to adjust its appearance. To address this, a fine-grained appearance translation strategy is designed in this work. When adjusting the appearance of each foreground region, our method pays more attention to the background regions that are more relevant to it based on content similarity and position information. Furthermore, a multi-scale feature calibration strategy is introduced to adaptively calibrate the fine-grained features. Finally, an adaptive reconstruction strategy is proposed to further improve the harmonization result. Extensive experiments show our method significantly reduces parameters and achieves state-of-the-art performance compared with previous methods.
What problem does this paper attempt to address?