Abstract:  Existing image completion work is mostly based on missing regions with regular, small area or sufficient context information. When the area of the region to be completed is relatively large, the completion work tends to be blurred or distorted due to the lack of context information and the instability of the generative adversarial network (GAN) training. Especially, if the missing area is located at the edge of the image, the final completion result will have large blank area or pseudo color. To solve the above two problems, the method of hierarchical joint image completion on the basis of GAN is proposed, and the network structure is improved to address the problem of unstable GAN training. On the one hand, it overcomes the problem of the generation of blank area in completion results due to the large missing area, thereby producing more realistic and clear texture details. On the other hand, it makes the adversarial network training more stable and suppresses the generation of pseudo-color. The experimental results demonstrate that the proposed method achieves better completion results.

Key words:  image completion, GAN, hierarchical joint, large missing area, edge missing area