Research On Artificial Image Completion And Translation Model Based On Wasserstein Generative Adversarial Network

There are four key tasks in the field of Computer Vision(CV): image detection,image recognition,image segmentation,and image generation.This paper mainly focuses on the image generation branch.With the gradual development of artificial intelligence to the cognitive stage,generation has become a key technology for the development of artificial intelligence.Image completion and translation have attracted a lot of attention.Image completion aims to restore the content of the lost area based on the known information in the image,so that the image remains true and reasonable both in the whole and local area.Image translation aims to transform the current image representation into another image representation through a deep neural network,which can be regarded as finding the mapping relationship between two or more image domains.At present,most algorithms are researched on natural images,and attention to special artificial images is relatively scarce.In order to make up for the research gap,this paper is based on the Wasserstein Generative Adversarial Network to carry out image restoration and image translation algorithm research on Chinese ancient paintings.The specific research content is as follows.Since the details of ancient Chinese paintings are very numerous and complex,and they have a large number of irregular textures,this paper proposes a detail-enhanced ancient painting completion model.This paper first analyzes the reasons why traditional image completion algorithms cannot solve the problem of artificial image completion,and designs a completion network with a multi-task branch structure for this problem.The ablation experiment proved that the branch structure can effectively improve the responsiveness of the model to the detail area,so that the repair result is more realistic and natural in detail.In addition,this article introduces a structural guidance mechanism,which allows users to participate in the process of completion.This paper verifies the module through user-guided experiments,and the results show that our model has a high degree of interactive ability.Finally,the whole network is constrained by a lot of loss,and extended experiments verify the versatility and robustness of our model.The traditional image translation model tends to gradually lose the underlying structural features of the image during the encoding process,so that it is impossible to generate a translation result with clear texture.In order to solve this problem,this paper introduces a variety of attention mechanisms combined with a multi-encoder architecture,and proposes a novel coloring model for ancient paintings.First of all,this paper designs a multi-encoder architecture to encode the line features of the ancient paintings and the color features of the reference image.The multi-encoder architecture helps reduce the difficulty of feature fusion,thereby improving the coloring effect.Secondly,this paper proposes a new channel attention mechanism in the content encoder,which enables the content encoder to effectively combine the color features with the most responsiveness of line features,thereby avoiding the influence of invalid information on the coloring process.Finally,this paper uses a skip connection mechanism with gated attention guidance to assist the content encoder to pass the underlying salient features to the decoder.A lot of experiments have shown that our model can reasonably integrate the color information of the reference image and the line information of the ancient painting.,and can effectively retain the underlying invariant features of the line image.In a word,our model effectively solves the problem that existing model does not work well on special artificial image data sets.