Human Pose Estimation Based on Ehanced Lightweight High-Resolution Network

Lightweight human pose estimation networks are designed to operate on mobile devices with limited resources and computing power. However, reducing model complexity often results in a decrease in accuracy. Addressing this challenge, this paper proposes a lightweight high-resolution network, Elite-HRNet, based on Lite-HRNet. To capture richer feature information, we propose an Information Enhancement Module (IEA), which includes the Information Enhancement Convolution (IEConv) and the Polarized Self-Attention Mechanism (PSA). IEConv employs a re-parameterization technique to convert multiple difference convolutions into ordinary convolutions, integrating prior information during the training process to enhance the model's capacity to extract differential information. PSA is a self-attention mechanism that enhances the long-range dependencies of feature information by integrating IEConv with PSA. Additionally, we incorporate Split-Concat-Shuffle (SCS) Modules into each branch of the network to further extract contextual information. Experiments demonstrate that the network proposed in this paper, when compared to Lite-HRNet, achieves an average accuracy improvement of 1.2% on the MPII dataset and an accuracy improvement ranging from 1.0% to 1.9% on the COCO dataset. While maintaining lightweightness, the network's accuracy is further improved.