Iterative facial image inpainting based on an encoder-generator architecture

From coarse to fine: Two-stage deep residual attention generative adversarial network for repair of iris textures obscured by eyelids and eyelashes

We propose a two-stage deep residual attention generative adversarial network (TSDRA-GAN) for inpainting iris textures obscured by eyelids. This two-stage generation approach ensures that the semantic and texture information of the generated images is preserved. In the second stage of the fine network, a modified residual block (MRB) is used to further extract features and mitigate the performance degradation caused by the deepening of the network, thus following the concept of using a residual structure as a component of the encoder. In addition, for the skip connection part of this phase, we propose a dual-attention computing connection (DACC) to computationally fuse the features of the encoder and decoder in both directions to achieve more effective information fusion for iris inpainting tasks. Under completely fair and equal experimental conditions, it is shown that the method presented in this paper can effectively restore original iris images and improve recognition accuracy.

SR-DeblurUGAN: An End-to-End Super-Resolution and Deblurring Model with High Performance

In this paper, we consider the difference in the abstraction level of features extracted by different perceptual layers and use a weighted perceptual loss-based generative adversarial network to deblur the UAV images, which removes the blur and restores the texture details of the images well. The perceptual loss is used as an objective evaluation index for training process monitoring and model selection, which eliminates the need for extensive manual comparison of the deblurring effect and facilitates model selection. The UNet jump connection structure facilitates the transfer of features across layers in the network, reduces the learning difficulty of the generator, and improves the stability of adversarial training.