仿真图像与遥感影像混合的训练方法研究  

Research on Training Method of Hybrid Simulation Image and Remote Sensing Image

作　　者：曹亚明 肖奇 杨震 CAO Ya-ming;XIAO Qi;YANG Zhen(National Space Science Center,Chinese Academy of Sciences,Beijing 101401,China;University of Chinese Academy of Sciences,Beijing 100049,China)

机构地区：[1]中国科学院国家空间科学中心,北京101401 [2]中国科学院大学,北京100049

出　　处：《计算机仿真》2023年第1期213-217,222,共6页Computer Simulation

摘　　要：针对真实遥感数据对于神经网络模型的训练效果有限、真实样本的获取在某些条件下受限、真实遥感数据所含特征有限等问题,提出将不同数量、分辨率的仿真图像作为部分训练样本,与真实训练样本相混合、组成新的训练数据集的方法。仿真图像与真实图像都可以单独用作训练样本,将两种训练样本混合后即形成了包含仿真图像与真实遥感图像的混合数据集,使用混合数据集对网络训练,并对混合数据集的样本均衡性进行研究。实验结果表明仿真图像作为训练样本可以有效提高网络对于遥感影像小目标的检测准确率,尤其是在有云雾遮挡的情况下。从混合数据集的均衡性考虑,随着仿真图像数量的增加,模型的准确率先提高后降低,不同分辨率条件下用作混合数据集的仿真图像的最佳数量不同。Aiming at the problems of limited training effect of real remote sensing data on neural network models, limited acquisition of real samples under certain conditions, and limited features contained in real remote sensing data, this paper proposes a new training data set method by mixing different numbers and resolutions of simulation images as part of training samples with real training samples. The experimental results show that the simulation image as a training sample can effectively improve the accuracy of the network for small target detection in remote sensing images, especially in the case of cloud and fog occlusion. Considering the balance of mixed samples, with the increase of the number of simulation images, the accuracy of the model first improves and then decreases. The optimal number of simulation images used as mixed samples is different under different resolution conditions.

关 键 词：仿真图像 训练样本 遥感小目标 神经网络 

分 类 号：TP391.9[自动化与计算机技术—计算机应用技术]