Multi-source material image optimized selection based multi-option composition
Hao Wu,Ding An,Xiaoyu Zhu,Zhiyi Zhang,Guodong Fan,Zhen Hua
DOI: https://doi.org/10.1016/j.imavis.2021.104123
IF: 3.86
2021-03-01
Image and Vision Computing
Abstract:<p>Image composition aims to composite a material region into a target image. Using this technique, more images could be interactive, as there are millions of images created daily in modern life. However, it is difficult for the majority of traditional composition methods to retrieve relatively realistic semantically valid material images. Also, even if the minority of traditional composition methods yield superior results, it remains difficult to retrieve adequate material images due to the limitation of existing semantically valid images. Artificial traces also significantly reduce the visual esthetic of composited results.</p><p>Based on these problems, we propose a multi-source material image optimized selection based multi-option composition method. Firstly, a robust sparse coding based retrieval model is used to retrieve material images effectively. On this basis, the attention mechanism based circular SA-GAN model could generate many semantically valid material images that guarantee the adequacy of material images. Secondly, the pixel-level optimization based multi-scale composition model minimizes artificial traces and reduces computational burden. Finally, an experimental database using many images was built. Based on this model, adequate comparative experiments using multiple evaluation criteria fully show the proposed method's effectiveness and robustness.</p>
computer science, artificial intelligence, theory & methods,engineering, electrical & electronic, software engineering,optics