| Human pose estimation is a technique for locating human body parts from input data（especially images and videos）captured by sensors.It is an important step for computers to understand human behaviors and actions,and it is also a crucial research direction in the field of computer vision.Human pose estimation provides rich human body geometry and motion information,which has been widely used in various task scenarios,including human-computer interaction,motion capture,augmented reality,etc.The emergence of deep learning makes human pose estimation methods more focused on improving the accuracy of pose estimation,ignoring key efficiency issues.Thus,the increasingly complex and large network structure brings huge number of parameters and computation,making it increasingly difficult to deploy to resource-constrained mobile devices.To solve this critical problem,model compression is gradually becoming one of the most concerned topics in academia.Based on a high-resolution human pose estimation network,this paper proposes a sparse model pruning method and designs a lightweight human pose estimation network.The main work and innovations are as follows:（1）This paper proposes a sparse pruning method based on a high-resolution human pose estimation network.First,L1 regularization is added in the training phase of the original model,and the weights of the convolution layers and the scaling factors of the batch normalization layers are sparsely trained to obtain a network structure with sparse weights.Then,the importance of the channels is evaluated according to the weights of the convolutional layers and the scaling factor of the batch normalization layers.Finally,the sparse pruning method is used to prune the unimportant channels.This method can reduce model parameters and calculations while maintaining model accuracy.The experimental results well demonstrate the effectiveness and advancement of this method.（2）This paper designs a lightweight high-resolution pose estimation network called SSENet（Shuffle Squeeze-Excitation Network）.SSENet is mainly composed of Shuffle unit and SSE unit.SSENet follows the design pattern of high-resolution pose estimation network,using Shuffle unit to replace the second 3×3 convolution in its first stage,using SSE unit to replace the residual unit in the remaining stage,and using depthwise separable convolution to replace the standard convolution in the multi-resolution fusion unit.SSENet shows better performance than popular lightweight networks on COCO dataset and MPII dataset,validating the effectiveness and superiority of this network. |
