Research On Image Inpainting Algorithm Based On Structure And Texture Information Guidance

Since the advent of generative adversarial networks,based on their powerful learning ability and discrimination ability,it has attracted the attention of many researchers.Subsequently,the researchers tried to apply the generative adversarial network to the pathological problem of image repair and achieved certain repair effects.However,with the increase of the masking rate in the image,accompanied by complex image texture and structural information,image inpainting algorithms based on generative adversarial networks generally appear to blurred restoration results,artifacts exist in some areas,and lost some feature information and other phenomena in the repair results.Based on this,based on the generative adversarial network,this thesis improves the network model from the perspective of improving the network’s ability to learn features and changing the convolution method in the network to improve the repair effect.The superiority of the algorithm proposed in this thesis is verified by experimental comparison with different image inpainting algorithms on the Celeba-HQ dataset and the Paris Street View dataset.The main work of this thesis is as follows:1.Aiming at problems such as blurring and artifacts in the restoration results,an image inpainting algorithm based on attention guidance is proposed.The generation network evolved from the U-Net variant.When the network extracts image feature information,a channel attention mechanism is introduced to obtain important structural feature information and texture feature information in the image.Then,a bi-directional gated feature fusion module is used to constrain the structure information and texture information and perform feature fusion,so that the overall consistency of the restoration results is maintained.To make the detailed information of the inpainted image more in line with human standards,while using the contextual feature aggregation module,the deformable convolution with the adaptive receptive field is chosen to replace the ordinary convolution,to make the inpainting result closer to the original image.By comparing with other algorithms on public data sets,it is proved that the restoration effect of this network is better than other algorithms,and the effectiveness of the proposed method is proved by ablation experiments.Experimental results show that the restoration results of the proposed algorithm are more realistic and in line with human visual standards.2.Aiming at the problem that the inpainting result is too smooth and some feature information is lost,an image inpainting algorithm based on recurrent feature inference is proposed.Firstly,the convolutional neural network is used to extract the texture features and structural features of the input image;then,by using the correlation between adjacent pixels,the extracted feature information is sent to the recurrent feature reasoning module,after repeated repetitions many times reasonable reasoning to reconstruct the feature map.The inference process is guided by a knowledge consistent attention mechanism to improve the overall details of the inpainted image.To enable the convolutional neural network to capture more feature information,we choose to use dilated convolution instead of ordinary convolution,because compared with ordinary convolution,dilated convolution has a larger receptive field.The results of comparative experiments with other algorithms prove that this algorithm has achieved good results in both subjective and objective evaluations.