Image Compositing Using Dominant Geometry Transformations of Patch Matching
Wu Ha · Jisuanji fangzhen · 2013
Blending images based on patch matching is an effect way to composite images, especially when the textures or structures of source image are expected to be aligned with that of target image in compositing region. A method to improve patch matching based image compositing was proposed by using dominant geometry transformations. It was observed that the geometry transformations(including shift, rotation and scale) used to match similar patches between different images were sparsely distributed, thus, the dominant transformations could be found to represent prominent patterns of transformations of patches. The object functions involved patch matching with these dominant transformations were optimized and the error of matching was reduced, which led to better compositing results. Moreover, limiting the search space to a few dominant transformations and local neighborhoods reduced the computational cost of patch matching.The experiments demonstrate that the method alleviates the blur in blending regions of compositing and aligns small misaligned texture better(the blur and the misaligned texture both caused by inaccurate patch matching). The performance of the method achieves up to 3x speedup compared to the image compositing method based on randomized patch search.