The Research Of Image Translation Model Based On Generative Adversarial Networks

In recent years,the field of computer vision has attracted more and more attention.As one of the important media of human communication,image has many important application scenarios in the real world,such as image recognition,image generation,image translation and so on.Among them,image translation is a research hotspot in this field.Similar to language translation,image translation refers to the transformation of images from features of one domain to those of another.With the development of deep learning technology in recent years,the research in the field of image translation includes image instance-level conversion,in addition to the initial image style transfer,black and white photo conversion,image super-resolution improvement,and face animation.Although existing studies have achieved image conversion at the level of style texture and resolution,there are still many problems in image translation at the instance level due to the difficulty in controlling the shape features of objects in the image.The conversion of object instances in images requires complex processing techniques and algorithms,so it is a very challenging task.In order to solve the difficulty of this task,this dissertation focuses on the image-to-image translation model based on generative adversarial network,studies its algorithm and improves the model,so as to improve the translation effect of image translation at the instance level.The detailed research content of this dissertation is mainly divided into the following points:Firstly,a mask-guided image translation model is proposed to solve the problem of object shape feature loss in the image.This generative model combines image segmentation mask,uses generative adversarial network and cyclic network to learn the encoding of background and posture of source domain and shape and style of target domain in potential space.After up-sampling and decoding,the encoding features learned by the generator are divided into two branches to generate effective sample original image and sample mask respectively.Both the generated sample and the original sample are fed into the discriminator,which determines whether it is true or false.Through confrontation,the image retaining the shape features of the objects in the foreground is finally generated.Secondly,a loss function based on feature similarity is proposed to solve the problem that the boundary between foreground and background is not clear in the conversion process.The loss function is mainly used to constrain the features of datasets such as posture,proportion and shape in the process of network training,so as to separate the foreground and background in the image and improve the image instance level translation effect.Finally,experiments were carried out on the MS-COCO dataset to verify the validity of the proposed network and loss functions.Compared with Cycle GAN and Insta GAN,both Inception Score and Fréchet Inception Score are improved.At the same time,due to the inadequacy and controversy of the existing evaluation indexes,we have added the experiment of human preference evaluation index,which has also achieved a great improvement in the evaluation of human preference.