Research On Extraction Of Human Skeleton From Images

The image-based human skeleton extraction is the basis for human pose estimation.The process is mainly divided into two parts: first,the individual joint points of the human body are extracted according to the information such as the texture and color of the human body in the image,and then the joint points are linked to form the human skeleton.Uncertainty of lighting environment,clothing posture and other factors will cause the final prediction accuracy to decrease.Therefore,most methods currently use multi-scale and multi-receptive field image representations for network training to improve the accuracy of joint detection by the network.However,while enriching the effective information representation,the information unrelated to the target task also increases,which becomes a bottleneck for improving network performance.At the same time,the simple feature fusion method fails to fully consider the global correlation between features at different scales,and cannot effectively fuse rich features.To solve the above problems,this paper uses the attention mechanism to capture the global dependence and focus on the important areas of the image.The combination of multi-scale features and attention mechanism is applied to the extraction of human skeleton,and the following research work has been carried out:1.The network structure of cascade pyramid network(CPN)is analyzed,the basic principle of attention mechanism is studied,and an improved attention-based multi-scale pyramid network(AMSPN)is proposed.Based on the analysis of the relationship between the network depth and the effective receptive field,AMSPN newly added a scale branche,which further enriched multi-scale information.A hierarchical attention feature enhancement module(HAFEM)was designed to enhance task-related information based on hierarchical attention while suppressing irrelevant information,breaking the performance bottleneck of CPN.2.In order to use information at different scales more effectively,a multi-scale balance module(MSB)is designed to balance features at different scales.The multi-scale balance module can dynamically update the attention distribution of features at different scales as the network is trained.Therefore,for specific joint point detection,the features of different scales can be more effectively used.At the same time,the multi-scale balance and difficult point mining mechanism are combined to furtherimprove the utilization of multi-scale information and the accuracy of difficult-to-detect joint points extraction.3.Aiming at the problem that the existing feature fusion methods fail to fully consider the global correlation between different scale features,the self-attention idea is introduced,and a multi-scale feature fusion method(RSMFF)based on regional similarity is proposed.The convolutional neural network itself has limited receptive fields,and the self-attention mechanism can easily capture global dependencies.RSMFF combines the self-attention mechanism and uses the similarity between global regions to perform multi-scale feature fusion,which makes the fused features have richer image information expression capabilities and can more effectively assist the network in skeleton extraction.Through the ablation experiments and comparative analysis on the MPII and COCO datasets,the effectiveness of the above method is verified.