Research On Image Inpainting Algorithm Based On Generative Adversarial Network

Image inpainting/completion is to fill the damaged area reasonably with the remaining information in the image.Its purpose is to realize the image filling and obtain the contact visual effect.As an important research topic in computer vision,image inpainting has been widely used in target object removal,human face replacement,old photo restoration,special effects,and so on.This thesis introduces the basic theory of the convolutional neural network(CNN)and generative adversarial network(GAN),as well as several basic image inpainting network structures based on deep learning.In addition,two image inpainting network algorithms are proposed on the basis of GAN framework.The main works in the thesis are summarized as follows.Firstly,this thesis proposes an image inpainting network model based on multi-stage inpainting network and multi-scale image discrimination.The generator network in the proposed model is a coarse-to-fine two-stage network.According to the structural characteristics of the generator,a two-stage loss function is designed in this thesis.By combining the convolution kernel with different parameters,the network implements the inpainting process from structure to detail.In addition,the discriminator network in our model adopts the combination structure of global and local discriminators,which can combine the multi-scale image features to obtain more realistic inpainted structures and details.The experimental results on three standard international datasets show that the proposed inpainting model can achieve better visual effects and higher quantitative evaluation results.Secondly,an image inpainting network model based on serialized attention module and mask prediction discrimination is proposed in this thesis.In the proposed model,a serialized attention module is designed between the encoding and decoding networks to disperse and arrange feature maps,and combine them into new feature sequences.After that,we use the attention mechanism in natural language processing to process the feature sequences,reconstruct them into the feature maps,and then input them to the decoding network in the generator network.Thus we can improve the ability of the generator network in extracting the long distance feature information of images.In addition,the discriminator network in our model adopts the network based on mask prediction.Namely,the discriminator network predicts the filling area using the inpainted images,and then compares the predicted filling area with the actual damaged area to achieve discriminant result.In this way,the discriminator network can predict the authenticity of image from the global perspective.The experimental results on three standard international datasets show that the proposed inpainting model can achieve better visual effects.Besides that,our model achieves higher quantitative evaluation results compared to the latest image inpainting approaches.