Lightweight and Efficient High-Resolution Network for Human Pose Estimation

To address the challenges of high parameter quantities and elevated computational demands in high-resolution network, which limit their application on devices with constrained computational resources, we propose a lightweight and efficient high-resolution network, LE-HRNet. Firstly, we designs a lightweight module, LEblock, to extract feature information. LEblock leverages the Ghost module to substantially decrease the number of model parameters. Based on this, to effectively recognize human keypoints, we designed a Multi-Scale Coordinate Attention Mechanism (MCAM). MCAM enhances the model's perception of details and contextual information by integrating multi-scale features and coordinate information, improving the detection capability for human keypoints. Additionally, we designs a Cross-Resolution Multi-Scale Feature Fusion Module (CMFFM). By optimizing the upsampling and downsampling processes, CMFFM further reduces the number of model parameters while enhancing the extraction of cross-branch channel features and spatial features to ensure the model's performance. The proposed model's experimental results demonstrate accuracies of 69.3% on the COCO dataset and 88.7% on the MPII dataset, with a parameter count of only 5.4M, substantially decreasing the number of model parameters while preserving its performance.

@article{Liu2024,
  title     = {Lightweight and Efficient High-Resolution Network for Human Pose Estimation},
  journal   = {International Journal of Advanced Computer Science and Applications},
  volume    = {15},
  number    = {8},
  year      = {2024},
  publisher = {The Science and Information Organization},
  author    = {Jiarui Liu and Xiugang Gong and Qun Guo},
  doi       = {10.14569/IJACSA.2024.0150824},
  url       = {https://doi.org/10.14569/IJACSA.2024.0150824}
}