2D Hand Pose Estimate Based On GCN Feature Boosting

With the continuous development of artificial intelligence and the popularity of deep learning,more and more artificial intelligence algorithms are applied to real life,which brings convenience to people's life.Hand pose estimation is a branch of artificial intelligence research in recent years.Hand pose estimation aims to generate 2D or 3D hand joint positions using pictures.Hand pose estimation can be used in VR,AR,assisted instruction and other scenes,and the research of hand pose estimation plays an important role.In hand pose estimation,because of the convenience of getting RGB images,many scholars are doing research on hand pose estimation based on RGB images.However,in the past research,most scholars ignored the context between the joints of the hand or only used the implicit context between the joints.At the same time,the use of the hand context can affect the accuracy of hand posture estimation.The implicit context relationship between joints is invisible to users,and it is completely realized by training,which is an uncontrollable relationship.The hand joints can form a natural relationship diagram,which can show the explicit relationship between the joints.However,in the research of hand pose hand pose estimation,no one uses the relationship diagram between the hand joints to estimate hand pose.Therefore,this paper aims to build a hand pose estimation network based on RGB images,which can consider the relationship between hand joints,and can use the relationship between hand joints to enhance features,while taking into account the differences between different hand joints.The main innovations of this paper are as follows:(1)according to the relationship between different joints of the hand,the characteristic relationship graph is constructed,which can effectively promote the flow of information between joints,and also fully consider the real structure of the hand.The characteristic relationship graph fully considers the spatial and symmetric relationship between the joints of the hand;(2))Using graph convolution neural network in stacked On the basis of hourglass network,a feature enhancement network Graph FB is constructed,which uses the hand joint feature graph to enhance the feature;(3)considering that when the graph convolution neural network aggregates the neighbor information,the neighbor information may have different effects on the current node,and each neighbor's feature has different effects on the current node,on this basis,it is innovative An information flow control gate is proposed to control the inflow of current node information from the characteristic graphs of different nodes,so as to avoid information redundancy and expand the inflow of useful node information at the same time.The experimental part of this paper compares three benchmark algorithms with two different public datasets,which shows the effectiveness of the proposed algorithm and the superiority of the comparison benchmark algorithm.