Image harmonization with spatial feature interaction and back-projection upsample

Abstract

Without any processing, the synthetic image visually unrealistic due to the established differences in the appearance of the foreground and background. In view of this situation, the task of image harmonization arises at the historic moment, and its purpose is to adjust the foreground appearance of a synthesized image to be closer to the background, thereby eliminating local visual differences. However, due to the limitation of the spatial feature interaction range in the feature extraction process, the global appearance transfer effect is not good. Therefore, to solve this problem, we propose an enhanced spatial feature interaction module. Meanwhile, we propose a back-projection up sampling module, which refines the reconstruction error during the reconstruction up sampling process and better restores the details of the reconstruction foreground. Our experiments on a public dataset, iHarmony4, show that the method effectively generates synthetic images with consistent overall appearance and enhanced detail.