基于二值神经网络的人体姿态估计方法

he human pose estimation task is one of the basic tasks in the field of computer vision and has been widely used in human-computer interaction, augmented reality, behavior analysis and other fields. Most of the existing pose estimation methods focus on improving the accuracy of the network, which leads to the design of more and more complex models, with little consideration of the deployment needs of the model. At present, most lightweight methods for human posture are pruning high-precision complex networks and designing lightweight network structures to reduce network computing resource consumption. However, the lightweight model obtained in this way still performs the multiplication of 32 floating point numbers. Addition operations still require high computing power. The weights and activations of the binary neural network are only represented by ±1, which consumes little computing power. Therefore, the main purpose of this paper is to study a human pose estimation algorithm based on binary neural network. Specifically, this paper introduces a scale factor to reduce the quantization error and improves the binarization function. In addition, the network is pruned to further reduce the computational cost of the network. Experiments show that the network proposed in this article achieves good results on MPII.