Research On Character Image Generation Algorithm Based On Posture Guidance

The core of the pose-guided person image generation model is a process that recreates the person image by replacing original pose with the expected pose for the same person image.It is a relatively challenging subject in the field of computer vision and faces many difficulties during the research.For example,the current person image generation models cause the composite person images with blurring and missing texture.In addition,some models even generate the person image with severely dislocated key points.Recently,with the development of human pose estimation technology,as well as the successful application of deep convolutional neural networks and generative adversarial networks in person image generation tasks,the person image generation work has achieved some significant results.However,it needs to be further explored.Therefore,the thesis puts forward a new research program to improve some problems in the current pose guided person image generation work on the grounds of the existing research,the main work of the paper includes the following:1)Due to the abstract semantic representation ability of convolution neural networks,this paper analyzes the correlation between the semantic contents and the image feature from the convolution neural networks to improve problems like the lack of appearance details in generated person images.2)Provide another generator structure model for the pose-guided person image generation by the blending and improvement of the deformable GAN model and the attention transfer network.This model uses a series of intermediate poses to gradually transfer the conditional person image's poses to the target person image's pose in order to increase the fidelity of the generated person image pose.As is indicated in the experiment,comparing with the previous experimental method,the improved one in this thesis raises the quality of generated person images.3)Propose a pose-guided person image generation model that incorporates feature feedback,which enhances the accuracy of pose integration in the first stage and the appearance of person images in the second stage by adding a feature feedback constraint mechanism in the two-stage of PG~2 network structure.At the same time,introduction of VAE in DCGAN makes it better to learn and retain the appearance semantic contents of conditional person images.Meanwhile,it makes the appearance details of synthesized person images be more delicate and true.The study shows that the person image generated by the pose guided image generation model,which incorporates features feedback in the article,is more similar to the target image and more in line with human visual perception.