Image Inpainting with Parallel Decoding Structure for Future Internet

• DOI: 10.3390/electronics12081872
• Publication Date: 2023-04-17T05:34:15Z
• Authors: Peng Zhao; Bowei Chen; Xunli Fan; Haipeng Chen; Yongxin Zhang
• Keywords: Inpainting; Discriminator; Feature (linguistics)

Image inpainting benefits much from the future Internet, but memory and computational cost in encoding image features deep learning methods poses great challenges to this field. In paper, we propose a parallel decoding structure based on GANs for inpainting, which comprises single network network. By adding diet extended-decoder path semantic (Diet-PEPSI) unit encoder network, can employ new rate-adaptive dilated convolutional layer share weights dynamically generate feature maps by given dilation rate, effectively decrease number of parameters. For composed rough paths paths, use an improved CAM reconstruction decoder that results smooth transition at border defective areas. discriminator, substitute local discriminator with region ensemble attack restraint only recovering square, like areas traditional robust training loss function. The experiments CelebA CelebA-HQ verify significance proposed method regarding both resource overhead recovery performance.