Joint compression and acceleration based on YOLOv3-MobilnetV1

Abstract:
Traditional object detection algorithms have higher space complexity and longer reasoning time. This problem limits the deployment of deep neural networks on resource-constrained mobile devices and embedded devices. In order to solve this problem, this paper studies several model compression techniques based on the YOLOv3-MobilnetV1 object detection algorithm. By using the COCO data set, we compared the differences in the applicability of model compression and fast reasoning between pruning, quantification, knowledge distillation and their combination of object detection algorithms. The experimental results show that the optimization model using multiple compression methods reasonably achieves a compression rate of nearly 10 times without errors.