Research On Pose Estimation Of Occluded Portraits Based On Deep Learning

The human pose estimation can be used to determine the presence position of the human body joint points,which can be used as a bridge between action recognition and human-computer interaction.The traditional pose estimation networks cannot accurately predict the joint points position under occlusion,and it is prone to incorrect joint points connections when crowded,and it cannot effectively deal with complex background information interference problems.The pose estimation methods based on deep learning make full use of the effective information and image features in the images,and use the superiority of neural networks in the field of object detection and image processing to obtain the correct position of human joint points.Therefore,it is very necessary to use deep learning methods to build a human pose estimation network that can effectively handle complex occlusion situations.Based on the above background,this thesis studies the pose estimation of occlusion portraits based on deep learning,the existing high-resolution pose estimation network is improved and optimized to design a new network--high-resolution feature generation recovery network.The work done is as follows:In order to ensure the resolution of the input images during deep learning network training,the backbone structure of high-resolution network is selected as the research basis to avoid the lack of feature map information due to insufficient resolution.The fusion attention mechanism is introduced to screen useful feature information channels and filter out background interference.The deconvolution and multi-scale feature fusion module are added to layer the pose estimation task of the small target portrait and the pose estimation task of the large and medium target portrait.Multi-scale feature information is obtained by integrating multi-scale feature maps at various levels.The generative and adversarial module is designed according to the idea of generative adversarial networks,the missing part is completed and predicted to obtain the correct joint points heatmap,and the final pose estimation result is obtained by optimizing the connection and optimal matching according to the joint points connection,and the output result map is visualized.To find out the true performance of the proposed improved network,a number of comparative experiments are conducted on the MSCOCO dataset and the Crowd Pose dataset,and the joint points prediction accuracy of each network is compared under different occlusion rate conditions.Based on the data of a number of experimental results and actual test results,the improved model proposed in this thesis outperforms other networks and has better performance in dealing with occlusion portraits in complex scenes,which proves the feasibility and superiority of using the ideas of highresolution networks and generative adversarial networks to deal with occlusion portrait pose estimation problems.