一种改进YOLOv4的遥感影像典型目标检测算法  

A Typical Target Detection Algorithm in Remote Sensing Images Based on Improved YOLOv4

作　　者：余培东 王鑫[1] 江刚武[1] 刘建辉[1] 徐佰祺 YU Peidong;WANG Xin;JIANG Gangwu;LIU Jianhui;XU Baiqi(Information Engineering University, Zhengzhou 450001, China)

机构地区：[1]信息工程大学,河南郑州450001

出　　处：《测绘科学技术学报》2021年第3期280-286,共7页Journal of Geomatics Science and Technology

摘　　要：如何使传统神经网络算法对遥感影像典型目标检测表现出良好的适应性是当前遥感领域的一项难点。在深入解析最新YOLOv4网络结构及算法核心思想前提下,首先通过增加104×104的特征层尺度并嵌入SE模块进行网络结构改进;继而根据待检测目标尺度特点调整锚点框,提高YOLOv4算法对遥感影像典型目标检测性能;最后设计对照实验进行验证。实验结果表明YOLOv4算法相比RFB-Net和RetinaNet有明显的性能优势,所提出的YOLOv4改进算法对遥感影像中飞机和油罐两类典型目标的召回率和平均准确率得到显著提升,证明了改进算法的有效性。It is a difficult problem how to make traditional neural network algorithm show good adaptability to the typical target detection of remote sensing image in the field of remote sensing.In probing the latest YOLOv4 core idea,network structure and algorithm,the network structure is firstly improved by adding 104×104 feature layer scale and embedding SE module.Then,the anchor point frame is adjusted according to the scale characteristics of the target to be detected,and the typical target detection algorithm performance of YOLOv4 is improved for remote sensing image.Finally,it is verified by designing contrast experiment.Experimental results show that YOLOv4 has obvious performance advantages in comparison with RFB-NET and RetinaNet.The proposed improved YOLOv4 algorithm significantly improves the recall rate and average precision of the two typical targets of aircraft and oil tank in remote sensing images,which proves the effectiveness of the improved algorithm.

关 键 词：YOLOv4算法 遥感影像 目标检测 特征尺度优化 SE模块 

分 类 号：P237[天文地球—摄影测量与遥感]