Generative Adversarial Networks Based Image Inpainting Technology

Image inpainting is a technique for realizing the repair of the image defect portion by using an algorithm.Since digital image processing technology was introduced in 2000,it has been receiving attention so far.Many excellent algorithms are widely used in the fields of cultural relics and art restoration,film and television production,medical imaging,public security criminal investigation systems,and remote sensing image processing.At present,the existing image inpainting algorithms are mature for small-area defects,scratches and text occlusion,the problem of the large defect area,lack of semantic information,complex texture and structural information remain difficult problems to be overcome.In recent years,with the advancement of machine learning research,especially the emergence of generative adversarial networks,it provides a good technical means to solve these problems.In this paper,to solve several key issues in image restoration technology based on generative adversarial networks,we conducted some research.The main works include the following three aspects:(1)To overcome the problem of the loss function in generative adversarial networks may lead to the disappearance of gradient in network learning,an image inpainting algorithm based on least square deep convolution generative adversarial networks is proposed.Firstly,the algorithm introduces the least squares loss function into the deep convolution generative adversarial networks,which is used to replace the sigmoid cross entropy loss function,this method greatly alleviates the gradient disappearance problem in network training.Secondly,a new inpainitng adversarial loss has been proposed,that is,the discriminator network based on the restored image as input,combined with context loss and generative adversarial loss,jointly optimize the input coding of the generator network,and improve the image inpainting quality.The effectiveness and superiority of the algorithm are verified by comparison experiments.(2)To overcome the problems of the training samples with complex backgrounds and few training samples could result,poor inpainting effect,an image inpainting method based on dense convolution generative adversarial networks is proposed.The algorithm combines DenseNet and GAN,and uses dense convolution block of DenseNet to construct an auto-encoding generator network,then uses this network to extract image features better through feature reuse.At the same time,in order to solve the loss of information transmission between network layers,skip-connections is introduced in the encoder-decoder structure of the generator network.The network is optimized for joint reconstruction loss,adversarial loss and TV loss while training the network.The implementation results show that the algorithm not only avoids the training network falling into local optimum,but also improves the quality of the inpainted image.(3)To overcome the problem of low similarity of random area defect image restoration,an improved image inpainting method based on double discriminator is proposed.The algorithm consists of three networks: generator,global discriminator and local discriminator.The local and global discriminator networks are used to jointly optimize the parameters in the generator network,and the visual coherence of the inpainted area in the overall image is improved.In this algorithm,the smoothing layer of the full convolution generator network is also improved,we employed dilated convolution in dense convolution block,which increases the effective transmission of the feature information,thereby improving the local restoration effect.Moreover,in the process of network training,we propose an approach to make the generator network has preliminary inpainting capability by pre-training with the reconstruction loss,and then use the joint loss training to generator network and the two discriminator network,thereby reducing the network training time.Experiments show that the algorithm can extract semantic information better and improve the similarity of the inpainted image.