Person Pose Synthesis Based On Deep Learning

Person pose synthesis is the task based on some pose key points to generate the same person but different pose.It is also an important branch of related to human poses.In recent years,research has found that person pose generation technology has great application value in data set expansion,video prediction and video synthesis.Although the human body image synthesized by the traditional method is close to the target image in posture and appearance,the fit of the posture and the target image is not enough,the retention of image details is poor,and the gap between the generated image and the target image is large.Therefore,in view of the above problems,this thesis combines the deep learning theory design model to improve the pose fit and detail retention.The work of this is as follows:1.This thesis improves the traditional generative adversarial network and innovatively proposes a dual input generation model.A network path to separately encode the key points image of the target pose,and then the original network to form a dual-input generative network.The pose feature of target image combines the feature of original image during,and finally a new image approaching the posture of the target image is obtained.Adding a coding path for pose information is equivalent to improving the utilization rate of target pose information in the generation phase.Generator obtains more pose information can generate the image closer target pose,thereby improving the problem of mismatch in the generated image pose.The experiment conducted on the public data set,and the results show that this method proposed in this thesis helps to improve the quality of the generated image.2.In order to enrich the details of the generated images,this thesis introduces the content reconstruction network from the content similarity.The content reconstruction network reconstructs the content of the generated image,restricts the distance between the generated image and the target image,and narrows the gap between the images by narrowing the distance between the high-level features of the image,so that the generated image is closer to the target image in content.This method of optimizing the distance of the high-level feature space to narrow the differences between the images makes the details of the generated images more abundant.3.In order to improve the quality of the generated image and reduce the complexity of the network,this thesis introduces an attention mechanism.The attention mechanism is embedded in the residual network to deepen the generated network and improve the ability of network feature extraction,reducing the loss of features as the network deepens.Similar to the human visual attention mechanism,which has a unique processing method for unique signals,the image attention mechanism makes the network give greater weight to the posture-related information of the generated image,making the network more concerned about the posture of the generated image,and the posture Irrelevant information attention mechanism gives less weight.Because the attention mechanism is embedded in the network,using the attention mechanism reduces the size of the network,and the generated image is closer to the target image.Experiments on public data sets that this kind of network with attention mechanism can produces high quality images.