Research On Multi-stage Image Expansion Method Based On Generative Adversaria

Human beings have a natural ability to perceive the invisible things around them based on visual content.When an incomplete image appears,human beings are good at filling the gap and making a realistic explanation for the possible missing content.This paper mainly studies the image outpainting tasks which are widely used in the fields of computer photography,image editing,computer graphics and so on.Compared with image inpainting,which requires deep neural network to understand the semantic content of natural images in order to recover the lost areas in photos,image outpainting is more challenging in space.In fact,the image outpainting task is to expand the image content,not to fill in the interior of the image.We are mainly committed to inferring invisible content beyond the boundary of real images to produce richer and more real images.Although some achievements have been made in image outpainting,there are still some problems,such as semantic inconsistency,unclear structure,foreground and background independent pixel interference and so on.The two research contents of this paper aim to solve the above problems.Firstly,for the problem that the existing image generation model is not suitable for the spatial outpainting of the image,resulting in the loss of structural and semantic information,and finally directly resulting in the fuzzy structure and content distortion of the generated image,we are inspired by the idea that humans first draw the boundary contour,then draw the color,and finally fill the content,A three-stage countermeasure generation network model(ECPIO-Net)is proposed,which decomposes the image outpainting task into contour prediction network(EP-Net),color prediction network(CP-Net)and image outpainting network(IO-Net),taking the predicted edge map and color map as the prior knowledge of the subsequent generated images,constrains the shape and content from both structural and semantic information,so as to make the generated images more natural.In order to further enhance the anti generation ability of the network,we propose a general conditional training strategy CTS,which can be integrated into other Gan networks for Enhance network training ability.Finally,the PSNR index of our image extended network model on the image reaches 27.1538,which is 8.59% higher than the latest method sienet.Secondly,in order to solve the problem of image distortion and artifact caused by the mutual interference of background noise when expanding the image,we further propose an optimization method of foreground and background expansion and fusion based on segmentation mask(hereinafter referred to as IOF).In this IOF model,an image is used as the input,the foreground and background are separated and expanded,and the output is a complete extended image.The research of this method is divided into two parts.First,a matching model is trained as the main pre-extractor,and then two networks are trained in series,one as the main extension generation network moo net and the other as the background extension generation network Network BGO-Net.In the first segment of the subject outpainting network,the input image separates the subject map through the matching model to generate the prediction map against the output subject,which is fused with the input map as the input of the second segment network,and finally output as the expansion map of the complete image.The CTS proposed above is used in both networks to enhance the training ability of the network and improve the quality of the generated image.