LENet: A Lightweight and Efficient High-Resolution Network for Human Pose Estimation
Ming Zhang, Xiandong Yu, Wenqiang Li, Xin Shu, Lei Pan, Zhongwei Shen · IEEE Access · 2025
Due to the huge requirements of performing human pose estimation tasks on edge devices with limited resources, more and more researchers have turned to work on the design of lightweight human pose estimation networks. As a typical lightweight network, Lite-HRNet achieves high performance in human pose estimation with a relatively low model complexity, but its inference speed is not ideal in practical applications. In order to address this issue, we propose a lightweight and efficient network (LENet), which is not only capable of learning comprehensive information under a lightweight architecture for real-time human pose estimation, but also has a faster inference speed than other common lightweight networks, so that the decision-making efficiency and user experience of real-time systems would be improved. We design two blocks, Recursive Fusion Block (RFB) and Deep Shuffle Block (DSB), to construct the model architecture. The RFB implements multi-scale feature fusion in a more lightweight way, it only carries out scale transformation between adjacent branches. The DSB fully utilizes computational resources to extract expressive information during the whole processing. Experimental results demonstrate that with the same complexity as of Lite-HRNet, our LENet yields inference speeds of 40.8 FPS and 150.4 FPS on CPU and GPU respectively, obtaining a great improvement of 83% over that of Lite-HRNet. Furthermore, LENet is more competitive than Lite-HRNet in terms of achieving a superior tradeoff among inference speed, overall performance and complexity of the model.