Image harmonization system based on self-attention transformation

The invention relates to the technical field of image processing, and particularly discloses two non-decoupling and decoupling image harmonization systems based on self-attention transformation, which utilize the strong remote context modeling capability of a self-attention transformation network and adopt a non-decoupling image harmonization module, so as to guide the harmony of the synthetic image through fully mining the relationship between the foreground and the background in the feature space of the synthetic image by using an auto-attention transformation network. or, a decoupling image harmonization module is adopted, and the implicit vector code of the background image light is decoupled by using an auto-attention transformation encoder and a decoder, so that the background light implicit vector code and the reflectivity feature image are fused through another auto-attention transformation decoder to generate an illumination intrinsic image; finally, the reflectivity intrinsic image is multiplied by