Abstract: Small celestial body (SCB) exploration is a research hotspot in deep space exploration technology. However, the scarcity of small celestial body surface images severely hinders the development of probe intelligence systems, and it is challenging to redeploy probes to obtain massive images due to the high cost. In recent years, with the growth of deep learning, generative adversarial networks have shown excellent potential in generating real images. This paper presents a high-quality small celestial body surface image synthesis method SCB-GAN, which can generate a massive amount and large diversity of small celestial body surface images that match human perception from a limited dataset. Specifically, we propose MineAug, which can automatically select appropriate data augmentations from the given SCB surface imageset based on an efficient mining mechanism, improving the robustness of the GAN training process. We also add the vision-aided loss to our network to further improve image quality. Through extensive experiments, our method outperforms recent methods under a variety of metrics in terms of visual quality comparison, no-referenced image quality assessment, and human subjective evaluation, which has significant implications for the development of intelligent systems.

The two datasets (Rosetta2495.zip and Dawn2500.zip) are preprocessed by dataset_tool.py and can be used directly for training SCB-GAN.

(This step is unnecessary since you can directly refer to the Information and proceed with the Step 2 directly.)

Run Execute DA.ipynb and Select DA.ipynb in sequence. The best K data augmentations selected by MineAug for Rosetta2495 and Dawn2500 are shown in the table above.

You can use the following command to train SCB-GAN. The training took about 5days on four Nvidia Tesla V100 (16GB) GPUs with a batch size of 16.

# train SCB-GAN with Rosetta2495 (Fig.6)
 python vision-aided-gan.py --outdir /outdir --data /Rosetta2495.zip --cfg paper256_2fmap  --aug ada --augpipe bgc --augcv ada --batch 16 --gpus 4 
 --kimgs-list '4000,2000,2000'  --num 3 --mineaug --augmentations-list 'HorizontalFlip,RandomAffine,Rotate'

 # train SCB-GAN with Dawn2500 (Fig.7)
python vision-aided-gan.py --outdir /hy-tmp/models/ --data /Dawn2500.zip --cfg paper256_2fmap  --aug ada --augpipe bgc --augcv ada --batch 16 --gpus 4 
--kimgs-list '4000,2000,2000'  --num 3 --mineaug --augmentations-list 'HorizontalFlip,Translation_y,Zoom_in'

# calculate metrics of SCB-GAN trained with Rosetta2495 （Table IV）
python calc_metrics.py --metrics=fid50k_full,kid50k_full,pr50k3_full --data=/Rosetta2495.zip --network=/network-snapshot-best.pkl

# calculate metrics of SCB-GAN trained with Dawn2500（Table IV）
python calc_metrics.py --metrics=fid50k_full,kid50k_full,pr50k3_full --data=/Dawn2500.zip --network=/network-snapshot-best.pkl

We have generated 5,000 images for each of the two datasets (Rosetta2495 and Dawn2500) using the pretrained SCB-GAN, which will provide a sufficient source of high-quality SCB surface images for future research related to the deep space field.

We express our gratitude to Jingmin Zhang for her invaluable assistance in creating the figures, as well as to the ten researchers who aided us in conducting the human subjective evaluation. Additionally, we would like to extend our heartfelt appreciation to the associate editor and reviewers for generously devoting their valuable time and effort to meticulously review our manuscript. Our codebase is built on Vision-aided GAN and StyleGAN2-ada.