Research on Object Detection Network Based on Knowledge Distillation

Abstract

Object detection is an important technology in the field of computer vision, but a large number of model parameters make it difficult to deploy in embedded devices. Knowledge distillation can compress network models and improve network accuracy. However, this method is mainly used in classification networks whose output is category. This paper proposes an end-to-end knowledge distillation model and applies it to the object detection network structure whose output is a heatmap. This method is not affected by the backbone network structure, and can directly extract the feature information of the network, so that the model can learn the dark knowledge in the teacher network more quickly. Aiming at the problem of insufficient dark knowledge in the single-teacher network, a multi-teacher self-adapting method is designed to integrate knowledge to improve the effect of the lightweight network. Finally, we experiment on the PASACL VOC data set to verify the effectiveness of the method in this paper.