Spacecraft Component Detection Method Based on Randomized Image Enhancement

Spacecraft component detection is the foundation and prerequisite for conducting close-range operations such as autonomous docking, capture, and maintenance of space debris. For those close-range operations, the key to successfully completing them is to autonomously and accurately detect target information. At present, the commonly used method for detecting spacecraft components is the image-matching algorithm, which uses manually designed features to match and detect target features and has high reliability. However, this approach relies on manually designed features and is difficult to apply in interactive scenarios with non-cooperative spatial targets. The detection method based on deep learning does not require manual parameter design and also has global feature extraction and semantic segmentation capabilities. However, the reliability of this algorithm highly depends on the quality of the image dataset. When the image data used for deep learning model training is insufficient, the accuracy of the detection results will be very low. However, due to the difficulty in obtaining in-orbit spacecraft images and the sensitivity of the data, there is currently no reliable spacecraft image dataset for detection, segmentation, and component recognition. In response to the above issues, the author proposes a randomized image enhancement method for spacecraft component detection. This method can automatically generate a large number of spacecraft composite images and corresponding component mask images in batches, without the need for manual annotation processes. Firstly, design a spacecraft synthetic image generator for neural network model training, which can automatically generate spacecraft synthetic images and corresponding component pixel masks. Secondly, to address issues such as insufficient model generalization ability and complex real image noise, data augmentation is performed on synthesized image data based on domain randomization. By randomly adding background interference and noise, the synthesized image features can be more diverse. Then, a semantic segmentation network is designed based on Mask R-CNN, which can automatically and accurately detect the solar wings and antennas of spacecraft through training using synthesized image datasets. Finally, the effectiveness and feasibility of this method were verified through simulation testing using synthesized images of unknown spacecraft models and real captured images of in-orbit spacecraft.