基于小样本数据增强的航天器表面损伤智能检测方法

在轨运行的航天器表面形成损伤有可能导致严重的后果，需要对航天器进行在轨实时损伤检测。针对航天器损伤检测图像样本难以获取的问题，本文采用智能化检测方法，提出了一种用于航天器表面损伤样本扩充的生成对抗网络，该网络能够学习单张输入图像的特征纹理表示，从而生成大量与输入图像特征相似的细粒度尺度样本，实现了少量图像数据样本的扩充。利用YOLO目标检测算法在扩充的图像样本中进行表面缺陷与损伤的检测识别，获取了较高的检测精度，为未来航天器健康状态监测与评估、通用化服务机器人应用及太空原位建设等提供了技术支撑。

An Intelligent Inspection Method for Spacecraft Surface Damage Based on Small Sample Data Augmentation

The surface damage to a spacecraft in orbit may have serious consequences， and thus real-time damage inspection is required. In order to solve the problem that spacecraft damage image samples are difficult to obtain， in the paper， a generative adversarial network (GAN) for spacecraft surface damage based on small sample data augmentation is proposed by means of the intelligent inspection method. The network can learn the feature texture representation of a single input image， and generate a large number of fine-grained samples similar to the features of the input image， thus realizing the expansion of a small number of image data samples. The YOLO object inspection algorithm is used to inspect and identify the surface defects and damage in the expanded image samples， and high inspection precision is obtained. The proposed network can provide technical support for the future spacecraft health monitoring and evaluation， the application of generalized service robots， and the in-situ construction of space.