Multiscale Distance Attention Network for Texture Image Inpainting
With the development of image inpainting technology,the existing image inpainting methods perform well on flat images,but the inpainting effect on complex textured images is not good.In order to solve this problem,thanks to the powerful ability of convolutional neural network to process texture information,this paper proposes a multi-scale distance attention image inpainting network(MDAN),which constructs a symmetric attention structure to generate suitable character-istics.The interactive attention mechanism is used to connect the heads of multi-head attention,and the distance prior of multi-distance fusion is introduced.In the process of feature matching,not only the similarity of features is considered,but also the influence of the distance between features is considered.Experiments are carried out on the public dataset DTD,and the effect of the MDAN model is better than the current mainstream method.
万方数据
维普
