Image Inpainting Based On Deep Convolutional Generative Adversarial Network

In the past few years,with the continuous evolution of deep learning and the improvement of computing power,computer vision has made great and comprehensive development.In the task of computer vision,the problem of semantic based image inpainting is one of the most discussed but difficult one.Conventional image inpainting methods only take the structural similarities between the completed image and the original image as the focus of the inpainting,and have shown some initial success;however,in the face of large missing area in image inpainting,these methods are powerless.In recent years,with the development of deep learning,the method based on the depth models,especially the depth generative models,has achieved better results because it can make full use of the semantic information of the image to be completed and the original ones.Although compared with the traditional method,the image inpainting algorithm based on the depth models can do better,there are still some unclear or even failed cases.In view of this,this paper proposes two image inpainting algorithms based on the depth convolution generation confrontation model and improves the original method.The details are as follows: 1.An image inpainting algorithm based on deep convolution generative adversarial network is designed.In order to make the completed image closer to the original image,we reset the inpainting task as the conditioned image generation task,that is,learning the distribution of training data in the training stage,using the trained deep convolution generative adversarial network to generate the countermeasure network in the inpainting,looking for the image feature embedding which is closest to the image to be completed on the data manifold;at the same time,we designed the content loss and reconstruction loss to help finding the closest manifold.In addition,we also designed the consistency loss based on the deep convolution neural network to constrain the consistency of the generated image and the original image in the deep feature space.This method effectively improves the details and authenticity of the image.2.An image incomplete algorithm based single generator and double discriminators is proposed.In order to make the part to be completed closer to the details of the original image,we add a patch discriminator for the area to be completed to further refine the details of this part.Based on the above model,we add a patch discriminator structure.In addition,we also introduce neighbor loss,which makes the completed image similar to the original image in pixel level and structure level,so as to improve the semantic and detail information of the completed images and improve the quality of the completed images effectively.