Résumé: Image outpainting aims at extending the field of an existing image. A current challenge with image outpainting is that background noise may interfere with the expansion of the subject, resulting in visual distortions and artifacts. In this study, a subject-aware image outpainting (SAIO) method, which reduces background pixel interference and emphasizes the subject, is proposed to solve this issue. After training a Matting model as a pre-extractor for the subject, two networks are trained in series: the subject outpainting network (SO-Net) for subject extension and background completion network (BC-Net) for background extension. First, the Matting model is used to simultaneously extract the input image subject and separate the background, and the subject is transferred to SO-Net to generate the predicted subject. Second, the predicted subject is fused with the background separated in the previous step as the input of the second network. Finally, BC-Net outputs the complete image. To improve the training ability of the network and quality of the output image, both networks adopt the conditional training strategy. The qualitative and quantitative results show that our method achieves good performance; a performance of 28.99 in peak signal-to-noise ratio (PSNR) on the test dataset was achieved. The proposed method can be widely applied to intelligent image processing systems.

Langue: Anglais