基于难易样本平衡因子的遥感图像目标检测  

Remote sensing image object detection based on the balance factor of hard and easy examples

作　　者：黄聪 杨垚 王华军[1,2] 李忠玉 赵金泉 马瑜 万军 HUANG Cong;YANG Yao;WANG Huajun;LI Zhongyu;ZHAO Jinquan;MA Yu;WAN Jun(Chengdu University of Technology,College of Geophysics,Ministry of Education,Chengdu 610059,China;Chengdu University of Technology,Key Laboratory of Earth Exploration and Information Technology,Ministry of Education,Chengdu 610059,China)

机构地区：[1]成都理工大学地球物理学院,成都610059 [2]成都理工大学地球勘探与信息技术教育部重点实验室,成都610059

出　　处：《物探化探计算技术》2022年第1期117-121,共5页Computing Techniques For Geophysical and Geochemical Exploration

基　　金：四川省人工智能重点实验室项目(2020RYJ02)。

摘　　要：遥感图像目标检测一直是遥感图像领域中的热点和难点问题,旨在分类和定位感兴趣目标。近年来卷积神经网络理论技术的快速发展,有效地解决了传统目标检测方法特征提取不足的问题。在公开的遥感数据集RSOD-Dataset上,基于YOLOv3算法模型进行目标检测,为了适应遥感图像中目标小、背景复杂等难分类样本的特点,在检测模型中引入难易样本平衡因子来改变不同类别的损失权重,从而进一步提高遥感图像检测精度。为了证明这里改进方法的有效性,进行了一组对比消融实验。实验结果表明,改进的算法比YOLOv3算法的平均检测精度提高了6%,尤其是对于背景复杂的立交桥类别,平均检测精度有了明显的提高。因此通过改进YOLOv3进一步平衡了简单样本和难例样本的损失权重,有效地提高了遥感图像目标检测精度。Image object detection of remote sensing has always been a hot and difficult problem in the field,which aims to classify and locate interested objects.In recent years,the rapid development of convolutional neural network theory and technology has used to effectively solve the problem of insufficient feature extraction in traditional object detection methods.The YOLOv3 was used to detect such object using the public remote sensing RSOD Dataset.In order to adapt to the characteristics of small object and complex background in remote sensing image,a balance factor of hard and easy examples is introduced to change the loss weight of different classes in the detection model,so as to further improve the detection accuracy of remote sensing image.To prove the effectiveness of the improved method,a group of comparative ablation experiments were carried out in this paper.The experimental results show that the average detection accuracy of the improved algorithm is 6%higher than that of YOLOv3 algorithm,especially for the overpass with complex background,showing that the detection accuracy has been significantly improved.Therefore,further balances the loss weight of easy and hard examples was used by improving YOLOv3,and it effectively improves the object detection accuracy of remote sensing image.

关 键 词：遥感图像 目标检测 卷积神经网络 损失函数 难易样本平衡 

分 类 号：P631.4[天文地球—地质矿产勘探]