Image compositing using dominant patch transformations

Patch-based synthesis can produce composites with smooth transition regions even though source images have inconsistent textures and structures, but it often suffers blur and small misaligned textures caused by inaccurate patch matching. We present a method to improve patch-based image compositing by using dominant geometric patch transformations (including patch offsets, rotations and scales). When searching for the nearest patches from matching sources, we observed that the patch transformations are sparsely distributed, and thus dominant transformations could be found from statistics of patch transformations to represent prominent patterns of patch matching. By combining dominant transformations with neighborhood searching, the accuracy of patching matching is improved. The computational cost also decreases as the patch search space is limited to a few dominant transformations and their neighborhoods. The experiments demonstrate that the improved patch matching alleviates blur and aligns small misaligned textures better in image compositing. In addition, the composite obtained by our method is consistent with the target image in color contrast. The running time of our method achieves up to 3x speedup compared to the approach based on the randomized patch searching.