Research On Image Inpainting Technology Based On Multi-Dimension Feature Fusion

As the existing image repair models can not fully use the complete region to predict the regional features of the missing part when the object features are seriously missing,it may cause a series of problems such as the discontinuity of the repair features and the blurred texture of the repair details.Consequently,this thesis puts forth two multi-scale feature fusion image repair models to enhance the image repair effect.This thesis proposes an image edge region repair model based on multi-scale feature fusion,which can enhance detail texture and enhance the definition of the repaired region,all while maintaining image consistency.This is due to the clear content blurring issue in the repair of edge regions with a large amount of data.This method uses two stages to generate the confrontation network structure.Firstly,starting from repairing the edge information of the image,parallel extended convolutional structures are added to the edge repair network to generate black and white images.At the same time,the multi-scale expansion convolution fusion structure is added to the image completion network,which increases the loss of feature content,expands the Receptive field and enriches the scale of feature extraction.Subsequently,the color picture is appended to the initial stage’s image,leading to the ultimate repair outcome.Experiments have revealed that the model proposed in this thesis has a superior repair effect than other models for large edge areas.Select PSV and Place2 with low difficulty in data set.The model can achieve good results under low mask rate.A model of multi-scale feature fusion image fine repair,based on common sense knowledge,is proposed secondly.The dynamic memory network is employed to combine the external and internal features of the incomplete image,thus producing an incomplete image optimization map.Secondly,a generation antagonism network with gradient penalty constraint is used to guide the generator to perform rough repair,and the incomplete image is optimized to obtain the rough repair image to be repaired.Finally,the method of correlation feature coherence is used to further optimize the rough repair map,and finally the fine repair result can be obtained.We use Place2,RUIE,and Underwater Target as data set.Comparing them to existing visual effects and objective data advantage repair models,the experimental results indicated that the model repair was slower than the first method due to the commonsense knowledge iteration,the texture structure was quite reasonable and the visual effect and objective data were superior to other models.