【摘要】 本文以树莓派为实际运行环境给出了一种轻量化算法EfficientNet-YOLOv3。该算法基于YOLOv3算法,以EfficientNet主干特征提取网络替换原有的Darknet53主干特征提取网络,并对EfficientNet特征提取网络的输入图片及后续特征图的分辨率从228×228扩张至416×416。实验证明,EfficientNet-YOLOv3算法的模型参数量较原YOLOv3算法减少了82.62%,在VOC数据集中的mAP较原YOLOv3算法提高了1.63%,在树莓派中运行时,对内存与CPU的占用分别减少了5%和15%,且运行速度可达到0.48fps,满足了嵌入式平台的运行要求。更多还原

【Abstract】 The paper takes the Raspberry PI as the actual running environment and gives a lightweight algorithm Efficientnet-YOLOv3.The algorithm based on YOLOv3 took EfficientNet’s feature extraction network to replace the original Darknet53 feature extraction network and the EfficientNet characteristic extraction network input image and the subsequent feature graph resolution from 228×228 expansion to 416×416.The experiment results proved that the Efficientnet-YOLOv3 reduced the number of model parameters by 82.62%compared with the original YOLOv3,and the mAP in the VOC data set improved by 1.63%compared with the original YOLOV3.When Efficientnet-YOLOv3 runs in Raspberry PI,it reduces the memory footprint by 5%and the CPU by 15%,respectively,and it run at 0.48 fps,which is efficient enough for an embedded platform.更多还原