Towards Statistical Consistency for Deep Image Inpainting

A method of improving the statistical consistency for deep inpainted images is proposed. Firstly, non-linear high-pass filtered residuals and a deep neural network are respectively used to extract intrinsic identity signal (IIS). It is found that there is IIS statistical inconsistency between the deep inpainted images and the pristine images, and such statistical inconsistency universally exists for different image sources and different deep inpainting algorithms. Secondly, a generative convolutional neural network is proposed to make the IIS statistics of inpainted regions be consistent with that of the pristine-regions, while maintaining high visual quality. Finally, by randomly perturbing some parameters of the generation network within a reasonable range, images with diverse patterns can be generated. It can reduce the detection accuracy of their sources being identified. The effectiveness of the proposed method is demonstrated by comparing the IIS statistics among pristine images, deep inpainted images and generated images, and evaluated by various kind of forensic detectors.