Research on YOLOv3 model compression strategy for UAV deployment
Fei Xu, Litao Huang, Xiaoyang Gao, Tingting Yu, Leyi Zhang
Abstract
UAVs are often limited by limited resources when performing flight tasks, especially the contradiction between storage resources and computing resources when the huge YOLOv3 model is deployed on the edge UAVs. In this paper, we tend to compress YOLOv3 model in different aspects to achieve load availability at the edge. In this paper, deep separable convolution is introduced to reduce the computation of the model. Then, PR regularization term is used as the regularization term of sparse training to better distinguish scaling factors, and then the hybrid pruning combining channel pruning and layer pruning is carried out on the model according to scaling factors, in order to reduce the number of model parameters and the amount of calculation. Finally, since the training data is a 32-bit floating point number, DoReFa-Net quantization method is used to quantify the model, so as to compress the storage capacity of the model. The experimental results show that the compression scheme proposed in this paper can effectively reduce the number of parameters by 97.5% and the calculation amount by 82.3%, and can maintain the original detection efficiency of UAVs.