Research And Realization Of Character Pose Transfer Generation Method Integrating Self-Attention Mechanism

Character posture transfer is the task of using the key points of human posture to guide the same character to generate multi-posture images.It is a new research hotspot in the field of artificial intelligence.At present,with the continuous deepening of computer vision research,character pose transfer technology is more widely used in the fields of online virtual fitting,video synthesis,and data set expansion.Although the existing methods can generate images under target conditions,there are still phenomena such as unreasonable posture,insufficient texture,and inconsistent generated image styles.Therefore,in response to the above problems,this article first comprehensively analyzes the research status and research methods of character pose transfer generation,and uses a two-stage model based on a generative confrontation network to conduct pose transfer research.Then,according to the network characteristics,task characteristics and some problems of the model Carry out optimization and propose improved strategies and methods.The specific research work is as follows:1.Generate image styles for constraints,enhance its posture fit and detail retention,build a content-aware network,calculate the difference between the generated image and the target image,check the gap between the two layers of features,and then feed back the generated network,To promote the generated image closer to the target image in style and structure.Use feature-level loss constraints to promote the generation of images to retain the original style,richer details,and more reasonable posture transfer.2.In order to further enhance the texture similarity between the generated image and the target image,the Markov discriminant model is used to replace the traditional discriminator,and the authenticity of the local receptive field of the generated image is judged to improve the discriminating ability of the discriminator and strengthen the high resolution of the generated image.Preserving the details,and then optimizing the discriminator model makes the generated image texture details richer and more realistic.3.On the basis of the above research,the self-attention mechanism is introduced to enhance the network's ability to learn features,enrich the generated image details,and then build a multi-posture character generation model that integrates the self-attention mechanism.By studying the principle of the self-attention mechanism and analyzing its application in the field of image generation that can be improved,the optimized self-attention mechanism is embedded in the residual network to deepen the generation network,improve the feature extraction ability,and deepen the discrimination at the same time The network improves the ability of image discrimination.Since the self-attention mechanism is similar to the human visual mechanism,it has a stronger ability to control the long-distance rationality of the generated image,and the learned features are richer,making the generated image closer to the target image,and the data is disclosed in the mainstream The experimental verification on the set,the experimental results show that the model can generate more realistic images.