| Human pose estimation is a hot research direction in the field of computer vision,and it has a wide range of applications in the fields of behavior recognition,human-computer interaction,pedestrian re-recognition and behavior prediction.In recent years,significant progress has been made in human pose estimation by using convolutional neural networks due to the development of deep learning in the field of human pose estimation.However,since convolution can only extract local features and cannot effectively extract long distance features and channel features,how to solve the problem of extracting long.distance features and channel features has become one of the key problems to achieve high precision human pose estimation.In order to solve the above problems,the attention mechanism is introduced in this paper,and this dissertation carries out researches from three aspects of channel attention,spatial attention and feature pyramid attention to optimize the feature learning ability of convolutional neural networks.The main work are summarized as follows:(1)In order to solve the problem that spatial information of feature maps is unable to effectively utilized when multi-resolution feature representations are directly fused in human pose estimation task,the multi-resolution human pose estimation network is proposed based on the High-Resolution Net(HRNet)for structural design,namely GCT-Nonlocal Net(GNNet),which combines both channel domain and spatial domain attention mechanism and contains improved exchange units,Gateneck module and Gateblock module.The exchange units are improved to extract more useful spatial information from the various feature representations by adding spatial attention mechanism before the multi-scale fusions,which make the information fusions between the different resolution representations better and result in the final high-resolution representation containing richer representation information.In addition,the Gateneck module and the Gateblock module are able to model channel relationships more explicitly to extract channel information more effectively by introducing channel attention mechanism.And the verification results on MS COCO VAL 2017 dataset show that the proposed method can effectively improve the fusion effect of multi-resolution feature representation.(2)To solve the problems of lack of learnable parameters in up-sampling and lack of deep feature extraction after fusion of high-resolution feature branches in multi-resolution networks,the adaptive upsample attention and the adaptive pyramid attention are proposed in this paper.Firstly,the up-sampling method without learning parameters and the problems of feature loss caused by it are analyzed,and based on that,the up-sampling method based on attention mechanism is proposed to optimize the feature loss caused in the process of up-sampling.Secondly,based on that feature pyramid can learn input feature map at multiple levels and combined with channel attention for the structural design,the adaptive pyramid attention is proposed to improve the extraction effect of high-resolution features.Finally,the adaptive pyramid attention is added to the high-resolution human pose estimation network and verified by experiments on the COCO 2017 dataset.In addition,the verification results show that the proposed method can improve effectively the feature loss in the process of up-sampling and the extraction effect of high-resolution branch features. |
