文档介绍：第 卷 第 期 集美大学学报 （自然科学版）
26 4 SUN Shidan , ＺHENG Ｊiaｃhun , ＺHAO Shijia , HUANG Yiqi
( 1. Sｃhool of Marine Information Engineering, Ｊimei University, Xiamen 361021, China;
2. Navigation Institute, Ｊimei University, Xiamen 361021, China)
Ａｂｓｔｒａｃｔ:
Aiming at the problem of automatiｃ deteｃtion of wearing helmets and masks at the same time in
the sｃene of ｃonstruｃtion site and dangerous area, an improved YOLOv3 algorithm is proposed to enhanｃe the
deteｃtion aｃｃuraｃy of helmets and masks. Firstly, the ｃlustering algorithm in network is optimized. The weighted
kernel K-means ｃlustering algorithm is used to analyze the dataset, so as to seleｃt the anｃhor box more suitable
for small targets deteｃtion and improve the average aｃｃuraｃy and speed of deteｃtion. Seｃondly, it optimizes the
Darknet ｃharaｃteristiｃ network layer in YOLO network. The extraｃted quadruple down sampling feature map is
up sampling onｃe. The double up sampling is fused with the previous double down sampling, and then it is
transmitted to the subsequent network together with quadruple down sampling, eightfold down sampling and six-
teen times down sampling to reduｃe the miss deteｃtion rate of small targets. Experimental results show that, the
average