Research on Lightweight High-resolution Network Human Pose Estimation Based on Self-attention

At present, when deep learning networks in the field of human pose estimation improve prediction accuracy, often accompany the improvement of network structure complexity, which brings about the improvement of network model parameters and computational complexity, making it difficult to deploy on devices with small computing power for practical application. On the basis of HRNet, this paper devises a lightweight human pose estimation network SAGNet that integrates the self-attention mechanism and Ghost Module, which introduces the self-attention module to get higher prediction accuracy in the transition process from the third to the fourth stage of HRNet, and replaces the standard convolution in HRNet with Ghost Module for network lightweight. The experimental results show that in the same experimental environment, on the coco 2017 validation set, SAGNet greatly reduces the amount of parameters and the complexity of computation compared with HRNet while maintaining the prediction accuracy at a similar level.