Photomosaic generation is a popular non-photorealistic rendering technique, where a single image is assembled from several smaller ones. Visual responses change depending on the proximity to the photomosaic, leading to many creative prospects for publicity and art. Synthesizing photomosaics typically requires very large image databases in order to produce pleasing results. Moreover, repetitions are allowed to occur which may locally bias the mosaic. This paper provides alternatives to prevent repetitions while still being robust enough to work with coarse image subsets. Three approaches were considered for the matching stage of photomosaics: a greedy-based procedural algorithm, simulated annealing and SoftAssign. It was found that the latter delivers adequate arrangements in cases where only a restricted number of images is available. This paper introduces a novel GPU-accelerated SoftAssign implementation that outperforms an optimized CPU implementation by a factor of 60 times in the tested hardware.
展开▼
