动态图融合的图像和谐化方法OA

Image harmonization is a key task in computer vision,which aims to make the foreground compatible with the background of the composite image.To solve the problems of insufficient utilization of background information and destroyed visual coherence of the foreground region in existing local region matching-based harmonization methods,a dynamic graph fusion network(DGFNet)is proposed.In this network,the designed graph fusion module can adaptively construct the matching fusion relationship between the foreground and the background,and between the foreground and the fore-ground according to different input images.Among them,the background branch establishes a dynamic connection between the foreground and the highly similar background area through dynamic threshold screening,and thereby realizes fine-grained adjustment.The foreground self-fusion branch constructs a dynamic connection of foreground-related features to realize the internal matching fusion of the foreground and improve the local visual consistency of the image.In addition,the global perception decoder equipped with DGFNet uses the global background information to further globally calibrate the foreground.Experiments show that DGFNet reaches the state-of-the-art on both iHarmony4 and ccHarmony public datasets.