Research And Application Of Image Inpainting Based On Deep Convolutional Generative Adversarial Network

The image is one of the important carriers of multimedia information and an important source of human access to external information.With the development of digital technology,digital image has been widely used due to its advantages such as easy storage and easy transmission.However,digital image is easily damaged in storage,transmission and other links,and then leads to image degradation.Therefore,it is of great practical significance to study effective digital image inpainting techniques.The digital image inpainting mainly refers to inferring the information content of the damaged area by using the information of the undamaged area,and then filling in the damaged area to restore the image.Digital image inpainting technology has been developed for nearly 20 years.Researchers have proposed many various methods based on different principles.The inpainting technology based on the image's own information,such as Criminisi,only uses the undamaged area of the damaged image to complete it.If there is no similar information found in the undamaged area of the image,the algorithm will not work.Another type of retrieval-based restoration technology is to select similar images from a large number of images on the Internet to fill in missing areas.However,the material on the Internet is too complicated.It's unable to find the required picture material in this way,and there is no guarantee that the image will looks logical.In order to solve the above problems,this paper studies the theory of deep convolutional generative adversarial network and image restoration,and proposes an image restoration method based on deep convolutional generative adversarial network.The main works done are listed below:(1)Design the deep convolutional generative adversarial network.After training the training set image,the network can generate similar images to the training set.In order to improve the quality of the generated image,batch normalization and Dropout optimization are introduced in this paper,and the least squares is used instead of crossentropy as the loss function.The image generation experiments based on the MNIST dataset and the CelebA face dataset show that the images generated by this method are highly similar to the training set.This work provides a large number of high quality candidate images for image inpainting.(2)In order to find the most suitable image for the repair task in the generated large number of images,this paper defines the loss function for the input of deep convolutional generative adversarial network.Then we can obtain the best image through the gradient descent algorithm,and using the image to complete the damaged image.The proposed method is used to repair the face image missing the center and left face based on the CelebA dataset.Compared with TV,Criminisi and other traditional methods,the experimental results show that the proposed method has achieved better repair effect.The validity of the proposed method is further verified by the lack of inpainting experiments based on the Chars74 K dataset center.