| In recent years,with the continuous development of artificial intelligence and the continuous advancement of image restoration technology,digital image restoration technology has been widely used in medical image processing,historical relics restoration,security monitoring and other fields.Image repair is mainly about extracting and selecting features from the background area,and then filling the missing or occluded areas with pixels.At present,many image repair algorithms have been proposed based on different image repair problems,but there are problems such as large missing areas and inability to retain the texture information of the original image,the emergence of generative adversarial networks provides a very effective technical solution for solving these problems,generative adversarial networks can automatically learn the characteristics of images,better deal with complex image repair problems,the network has a large amount of training data,the quality of the generated images is also better,and the repair results are more refined in terms of processing damaged images.Based on the improved generative adversarial network,this paper designs a facial and mural image restoration method for generative adversarial network,and the main research content is as follows:First,aiming at face image repair with complex semantics,the missing context connection is unnatural,and it is difficult to generate high-confidence repair results on detail textures,this paper improves an irregular face image repair model for generating adversarial networks.The model has a generative network,an attribute extractor and a discriminant network,and mainly designs the structure of the generative network,which is similar to the encoder-decoder structure.Among them,the first three layers of standard convolution are replaced with gated convolution to improve the input of effective pixels,and a contextual attention mechanism is introduced when extracting features,which helps the model learn the connection between missing area pixels and surrounding pixels.A property extractor VGG16 has been added,which is a deep convolutional neural network containing a large number of convolutional layers and a small number of fully connected layers.The image is repaired by the generation network and the repair image is obtained,and the attribute extractor VGG16 extracts the attributes of the real image,decodes the extracted attribute and the output of the encoder to obtain another repair image,and sends it to the discriminant network for judgment,and the weight is shared between the two generated images,and the final repair result is continuously trained,so as to achieve high-precision repair of complex image detail textures with large areas missing.Second,the above method has good results for face image repair,but the repair effect of mural image is poor,the color of the mural is different from the natural image,and the overall color of the repair image will be inconsistent.To solve this problem,this paper improves an irregular mural image restoration model based on double restoration network,and mainly designs the structure of two repair networks,first the structure generation network repairs the structure of the mural,and then the color correction network corrects the color of the mural.This approach improves the authenticity of the content generated in the restored area and also guarantees the color consistency and coherence of the restored murals.In order to prove the effectiveness of the model,this paper performs qualitative and quantitative experiments on the same dataset,and from the results of the repair,the two image repair models proposed in this paper have better results in the repair of problems such as unnatural context connections.From the perspective of specific indicators,the repair results of the proposed model are improved in three indicators: structural similarity,peak signal-to-noise ratio and perceived loss. |
