融入部分卷积和解耦全连接注意力的轻量级人体姿态估计

Lightweight Human Pose Estimation of Incorporating Partial Convolution and Decoupled Fully Connected Attention

轻量级人体姿态估计网络使得网络的参数量和计算量大大减少,使其能够在计算资源有限的设备上进行快速推理。如何在保持模型轻量化的同时提高人体姿态识别网络的性能是当前重要的研究课题。文章基于Dite-HRNet,提出融入部分卷积和解耦全连接注意力机制的LPFANet网络,将部分卷积与动态分离卷积相结合,构建了一个强化特征提取结构,同时使用了全局特征建模和密集特征建模进行特征再提取。在MPII数据集上测试,实验表明,与DiteHRNet相比,LPFANet在少量增加参数量和计算量的情况下,平均准确率提升了1.2%。文章网络在轻量化的同时有效提升了识别精确度。

Lightweight human pose estimation networks greatly reduce the number of parameters and computational resources,enabling fast inference on devices with limited computing resources.How to improving the performance of human pose recognition networks while keeping the model lightweight is currently an important research topic.Based on Dite�HRNet,this paper proposes LPFANet network which incorporates partial convolution and decoupled fully connected attention mechanism.It constructs a strong feature extraction structure by combining partial convolution with dynamic separable convolution,and uses both global feature modeling and dense feature modeling for feature re-extraction.It tests on the MPII dataset,and the experiments show that LPFANet improves the average accuracy by 1.2%compared to Dite-HRNet,with a small increase in the number of parameters and computational resources.The proposed network effectively improves the recognition accuracy while maintaining lightweightness.