Image Inpainting Based On Generative Adversarial Networks

Image Inpainting refers to the technique of restoring the lost parts of the image and reconstructing them based on background information.It refers to the process of filling in missing data in the designated area of the visual input.In the digital world,it refers to the application of complex algorithms to replace missing or damaged parts of image data.In the application of digital effect image inpainting,image coding and transmission,image inpainting has been widely studied.In recent years,the rise of deep learning-based image restoration has shown more powerful restoration capabilities,but its texture structure,semantic consistency and other aspects still need to be improved.Therefore,this paper proposes two image inpainting models based on deep learning.First,although the Semantic Image Inpainting model(SI)can achieve a certain quality restoration effect,there are some shortcomings in the details of the inpainting image.In response to this situation,a GAN(Generative Adversarial Networks)is designed: with a fully connected layer,the generator uses multiple layers that are repeated several times,combining convolution and sampling.the discriminator uses residual blocks,pooling layers,a fully connected layer,and a self-attention layer in the middle layer to improve the ability of discriminator.The fully connected and convolutional layers of the entire network use the Spectral Normalization to ensure the stability of the training process.Then,the trained model is applied to semantic image inpainting.The weighted context loss and prior loss are introduced in the inpainting process,so that the reconstructed image not only conforms to the semantic information but also conforms to the real image distribution as much as possible.By comparing the qualitative and quantitative analysis with the SI method,it is proved that the method proposed in this paper is more capable of repairing images that are consistent with visual perception and the overall semantics.Second,the image inpainting method based on deep learning requires training a large amount of data,which often requires several days of training time;and the model has a large number of parameters,which will occupy a large amount of physical storage space.Based on the Learning to Incorporate Structure Knowledge for Image Inpainting model(SKI),this paper simplifies the generative network: reduces the residual module and removes the two attention layers.The generator inputs the incomplete image and its gradient map,and both of them are trained together.In the decoding stage,the repaired gradient map is used to guide the restoration of the image.The generator uses hybrid loss and structure loss,and the discriminator uses the form of Patch GAN.The experiment compares the training speed,model size,etc.with the SKI method,which shows that the model has been simplified to a certain extent.The index evaluation and visual comparison with other inpainting models show that the simplified model still has a high level of repair ability.