PatchSwapper: A novel real-time single-image editing technique by region-swapping

•We design a special image-editing tool explicitly for seamlessly swapping regions (transformed or fixed) in a single image.•A novel graph algorithm is used to search for the best closed cut. A Poisson blending is applied to further reduce artifacts.•A CPU multi-thread implementation achieves real-time performance even for large patches (e.g. size=350*350).•User can draw arbitrary boundaries to indicate two regions to be entirely exchanged after swapping (please see our video).•A patch location recommendation algorithm is proposed to help user find out potential exchangeable regions existing in image. [Display omitted] While many image composition and synthesis techniques have been proposed, neither existing work nor professional image editing softwares such as Adobe Photoshop© provide a program explicitly for the task of realistically swapping regions within a single image. In this paper, we present an easy-to-use image-editing tool explicitly for that purpose, named as PatchSwapper. Users can simply determine the centre and radius of a pair of candidate regions or specify arbitrary borders by sketches; then, the swapping step is automatically and optimally executed. A graph-based approach is designed to find the optimal borders of two irregular regions to avoid generating visible seams. We use this approach to handle both non-transformed and transformed patches. For non-transformed cases, we not only achieve realtime performance with CPU multi-thread implementation, we also provide a location recommendation algorithm to help users find the appropriate exchangeable areas. For transformed patches, our method searches for the optimal transformation to generate interesting appearance changes on both the source and target locations. A Poisson color blending is performed onto the stitched patches. Overall the proposed approach is not only suitable for swapping image objects of any shapes but also for some other applications, e.g., image completion and composition. The experiments demonstrate that common images often contain potentially exchangeable areas and that large content variations can be obtained by simply swapping their locations.