摘要
针对航天领域深度学习的监督训练数据集难以完全模拟真实地球背景与目标纹理的问题,提出一种基于深度学习的数据增强方法。首先,基于梯度惩罚式Wasserstein生成对抗网络生成虚拟太空背景图像,通过学习原始图像中的地球背景信息并生成相似的虚假纹理,丰富地球纹理变化;同时,训练一种随机化风格迁移网络对数据集的目标表面纹理风格进行随机生成。该网络在保留空间目标结构和图像高维语义信息的基础上,提高目标纹理的多样性。最后,在光照一致性约束下对生成的地球背景和表面纹理风格化的目标进行融合,构建完整的增强图像数据集。对位姿估计网络进行仿真实验;结果表明,利用该方法构建的增强数据集相比原始训练集,网络的位姿计算精度得到显著提升。
A novel data augmentation approach is introduced to mitigate the challenge of insufficient realism in supervised training datasets for aerospace applications,where simulated Earth backgrounds and target textures often diverge from actual space scenarios.Initially,virtual space background images are synthesized using a modified WGAN-GP framework,which captures latent Earth background features from original imagery and generates synthetically varied textures to expand the diversity of terrestrial patterns.Concurrently,a stochastic style transfer network is developed to dynamically alter surface texture characteristics of dataset targets.This network maintains spatial target integrity and highlevel semantic consistency while introducing texture variability through adversarial style recombination.Finally,synthesized Earth backgrounds and stylized target surfaces are fused under illumination consistency constraints,ensuring photorealistic integration and yielding a comprehensive augmented dataset.Simulation experiments on pose estimation networks indicate that the augmented dataset constructed with this method significantly improves pose estimation accuracy compared to the original training set.
作者
张凡
张泽旭
宋卓
黄烨飞
袁萌萌
ZHANG Fan;ZHANG Zexu;SONG Zhuo;HUANG Yefei;YUAN Mengmeng(Deep Space Exploration Research Center,School of Astronautics,Harbin Institute of Technology,Harbin 150080,China)
出处
《宇航学报》
北大核心
2025年第7期1456-1466,共11页
Journal of Astronautics
基金
国家自然科学基金资助(U20B2001)。
关键词
空间目标
数据增强
生成对抗网络
风格迁移网络
深度学习
Space target
Data augumention
Generative adversarial network(GAN)
Style transfer network
Deep learning
