摘要
针对遥感图像目标检测的目标尺度小、分辨率过低的问题,提出残差对抗目标检测算法的遥感图像检测方法。通过残差对抗的方式对图像的特征信息进行重构,完成图像分辨率的提升。在Backbone主干网络提取图像特征信息的基础上由Neck结构将特征进行融合,最后由CIoU_Loss损失函数提高定位回归精度,提高模型性能。实验结果表明,与其他算法相比,在平均精确率、平均召回率、平均综合指标F1值、平均mAP值方面分别提高了8.15%,6.9%,7.15%,6.75%。所提算法在低分辨率遥感图像目标检测方面准确性较高,对遥感图像小目标检测效果较好。
Aiming at the problems of small object scale and low resolution in remote sensing image object detection,a remote sensing image detection method based on residual adversarial object detection algorithm is proposed in this paper.The feature information of the image is reconstructed through residual adversarialism,so as to realize the improvement of image resolution.Based on image feature information which is extracted by Backbone network,the features are fused by Neck structure.Finally,the CIoU_Loss function is designed to increase the regression accuracy,and improve model performance.Experimental results show that,compared with other algorithms,the mean precision,mean recall,mean F1-score and mean mAP value of this algorithm are improved by 8.15%,6.9%,7.15% and 6.75% respectively.The algorithm has high accuracy in object detection in low-resolution remote sensing images,and has good effect on small object detection in remote sensing images.
作者
黄玉玲
陶昕辰
朱涛
司俊文
吕昌东
吴迪
沈占锋
HUANG Yuling;TAO Xinchen;ZHU Tao;SI Junwen;LYU Changdong;WU Di;SHEN Zhanfeng(Soochow University,Suzhou 215000,China;Aerospace Information Research Institute Chinese Academy of Sciences,Beijing 100000,China)
出处
《电光与控制》
CSCD
北大核心
2023年第7期63-67,共5页
Electronics Optics & Control
基金
国家自然科学基金(41971375)
苏州大学“大学生创新创业训练计划”资助(202110285074K)。
关键词
遥感
目标检测
超分辨率
残差对抗
remote sensing
object detection
super-resolution
residual adversarialism
