DMLNet:Densely Connected and Multi-Scale Lightweight High-Resolution Human Pose Estimation Network

The current research on human pose estimation mainly focuses on improving model accuracy, while efficiency optimization in resource limited scenarios such as mobile devices or edge devices has not been fully explored. Specifically, how to maintain high accuracy under low parameter count (< 5M) and real-time inference (> 30FPS) conditions remains a problem.This paper introduces a densely connected and multi-scale lightweight high-resolution human pose estimation network, termed DMLNet, which demonstrates superior performance in comparison to existing lightweight networks.DMLNet introduces the Stem module, TMCneck module, and KTConv module based on HRNet(high-resolution network). The Stem module overcomes the limitations of the original HRNet modules, which could only extract basic information from feature maps, enabling the capture of partial details and deeper information from these maps. The TMCneck module, with its innovative network structure and incorporation of attention mechanisms, significantly enhances the model's accuracy. Meanwhile, the KTConv module achieves both model lightweighting and the extraction of feature information across multiple scales.Additionally, we have introduced dense connections and integrated a feature fusion strategy that combines RFCA attention to enhance the model's performance.We conducted extensive experiments on the COCO and MPII validation datasets, and the results significantly surpass those of existing networks.