Research on pruning algorithm of target detection model with YOLOv4
Huang Linglin, Qiang Li, Xianzhen He, Lin Maosong · 2020
With the development of deep convolutional networks in target detection, deep convolutional networks are developing in the direction of deeper parameters with greater amounts. In order to reduce the parameters and inference time of the target detection network model, a pruning algorithm combining channels and layers is proposed and used for pedestrian detection. Through L1 regularization of the channel scale factor, the channels of the convolutional layer become sparse. The channels and shortcut layer with less information are pruned, so that the model is compressed. Pruning the YOLOv4 model based on this principle. Experiments prove that sparse training is important for model pruning. Through comparative experiments, the pruned YOLOv4 model can better detect targets on the INRIA dataset. Compared with the original YOLOv4, although there is a small loss of accuracy, the volume, parameter amount and reasoning time of the pruned model are greatly reduced.