Semantic Validation Based Massive Image Composition

Abstract

Creative composited results could be achieved by composition technique using target and candidate images. However, with the advent of big data era, how to composite a huge database of images gathered from different sources has become one non-negligible challenge. Traditional methods have the obvious drawback by ignoring semantic validation of massive images. Even though some algorithms have considered this point, the accuracy of semantic matching is not realistic.Aiming at the problem above, we proposed one semantic validation based composition method. On the one hand, optimized VGG16 model was used for retrieving semantically valid candidate images which ensures the semantic validation between candidate images and target images. On the other hand, Poisson blending and some related algorithms contribute to achieve the final composited results, especially for the boundary. At last, a database concluding a large number of images was built, based on it, sufficient experiments indicate that our method could achieve realistic multi-option composited results, especially for considering semantic validation fully.