Hierarchical Joint Image Completion Method Based On Generative Adversarial Network

Image completion technology refers to the method to inpaint an image with missing or damaged area into a complete image based on the context information,and achieves a natural and true complement result.Image completion technology is a fundamental problem in low-level vision and has now attracted widespread interest in computer vision and graphics.Most of the existing image completion methods are based on the small area or the missing area in the center of the image.When the missing area is large or located at the edge of the image,the result will be blank,distorted and pseudo-colored.Based on the deep learning,this paper proposes a method of hierarchical joint image completion method on the existing methods,which is to complement the large area or the missing area at the edge of the image with generated adversarial network.Firstly,the mean square error loss between the complemented image and the original image is taken as the objective function,and the defective image is iterated through the completion network to obtain the preliminary completion result;then the completion network is fixed,and the result is divided into three parts,the whole image,the area centered on the missing area,and the area of the center of the missing area,and input them into the corresponding global discriminant network and the local discriminant network for adversarial training.In addition,the network structure has been improved for the instable training of the original generative adversarial network.The advantage of this method are: on the one hand,the method ensures the global and local consistency of the completion result and the realistically of the generated texture information by partitioning the image;on the other hand,it overcomes the problem of unstable training of the original adversarial generate network by adding a local discriminator and adjusting the original network structure,which depress the generation of pseudo-color when the missing area is large or locate at the edge of the image.The experimental results show that the proposed method has better results in reconstructing the image with large missing area or missing area at the edge of the image.The result of the completion is more real and natural.