Adaptive Visual Field Multi-scale Generative Adversarial Networks Image Inpainting Base on Coordinate-Attention

Image inpainting with the large missing blocks is tremendous challenging to achieve visual consistency and realistic effect. In this paper, an Adaptive Visual field Multi-scale Generative Adversarial Networks (denoted as GANs) Image Inpainting based on Coordinate-attention (denoted as AVMGC) is proposed. Firstly, an encoder with deformable convolutional networks in the generator of multi-scale generative adversarial networks is designed to expand the local vision field of network sampling adaptively in the image inpainting, which improves the local visual consistency of the image inpainting. Secondly, in order to expand the receptive field of the deep network and the global visual field, AVMGC combines the coordinate-attention mechanism with the convolutional layers, aiming to capture the direction-aware and position-sensitive information by cross-channel, which helps models to more accurately locate and recognize the objects of interest and generate globally consistent geometric contour in the image inpainting. In particular, instance normalization is introduced to the mutil-scale discriminator for transferring the statistic information of the feature maps and aims to keep the style of the original images. Extensive experiments conducted on public datasets prove that the proposal algorithms have the qualitative performance and outperform the baselines.