摘要
随着深度学习的快速发展,基于特征金字塔设计的检测器在光学遥感舰船检测领域取得了重大进展。然而,这些检测器在检测较小舰船目标时仍会出现漏检或误检的情况,因此针对光学遥感舰船小目标所存在的候选区域小、语义信息不匹配和特征信息不充足等特性,改进了特征金字塔网络(FPN)中特征图上采样的方法和其自上而下的融合网络,并通过控制相邻层级间的特征融合因子来丰富小目标的特征信息,提升浅层特征图对小目标特征的学习能力。实验结果表明,该方法能够有效提升检测器对舰船小目标的检测性能。
With the rapid development of deep learning,the detector based on feature pyramid design has made great progress in the field of optical remote sensing ship detection.However,these detectors will still miss or misdetect small ship targets.Therefore,the sampling method on the feature map and its top-down fusion network in feature pyramid network(FPN)are improved to address the characteristics of small candidate regions,semantic information mismatch and insufficient feature information in optical remote sensing small ship targets.By controlling the feature fusion factors between adjacent levels,the feature information of small targets is enriched,and the learning ability of shallow feature maps for small target features is improved.The experimental results show that this method can effectively improve the detection performance of detector to small ship targets.
作者
杨程
林泽南
王麒风
陈舒敏
林晨
YANG Cheng;LIN Zenan;WANG Qifeng;CHEN Shumin;LIN Chen(The 723 Institute of CSSC,Yangzhou 225101,China;Navy Equipment Department,Beijing 100084,China)
出处
《舰船电子对抗》
2023年第5期23-29,74,共8页
Shipboard Electronic Countermeasure
关键词
小目标检测
光学遥感
航空图像
卷积神经网络
small target detection
optical remote sensing
aerial image
convolutional neural network
