An optimization high-resolution network for human pose recognition based on attention mechanism

In the high-resolution network (HRNet), the low layer of low resolution part can adopt shallow parallel network structure to maintain the high-resolution features and highlight global features. However, the high-resolution human posture estimation network has the problems of large amount of network parameters, high complex calculation and low recognition precision of similar actions. To solve these problems, we proposed an optimized HRNet based on attention mechanism. Firstly, the dilated convolution (DC) module is introduced into cross-channel sampling to obtain global features by increasing the receptive field of the feature map, which ensures that the feature map can cover all the information of the original image; Secondly, the channel attention Squeeze-and-Excitation (SE) module is introduced in the process of cross-channel feature fusion to learn the correlations, which can recalibrate the features, highlight the information features selectively and suppress the secondary features, improving the recognition precision without changing the parameter quantity and operation complexity; Finally, the experiment results on KTH dataset show that the HRNet with channel attention mechanism and dilated convolution has better accuracy.