Learning Large-Factor EM Image Super-Resolution with Generative Priors (GPEMSR)

Accepted by CVPR-2024
Jiateng Shou, Zeyu Xiao, Shiyu Deng, Wei Huang, Peiyao Shi, Ruobing Zhang, Zhiwei Xiong, Feng Wu*
MoE Key Laboratory of Brain-inspired Intelligent Perception and Cognition, University of Science and Technology of China
Institute of Artificial Intelligence, Hefei Comprehensive National Science Center
Suzhou Institute of Biomedical Engineering and Technology, Chinese Academy of Sciences
*Corresponding Author

Code instruction

We provide comprehensive code implementations for the CREMI dataset. With minor modifications, the code can be readily adapted to the AC4 dataset. We provide all the pre-trained models.

Training

Three stages of our model are trained separately.
We provide all the training code and configuration files.

Testing

CREMI

Please put all the trained models in the 'pre-train_model' folder.

Output test images

python output_GPEMSR.py -opt option/output_GPEMSR_x8.yml # for 8x EMSR  
python output_GPEMSR.py -opt option/output_GPEMSR_x16.yml # for 16x EMSR

Segmentation metrics

Please put the pre-trained segmentation models in the 'pre-train_model' folder.

Superhuman	MALA
GoogleDrive	GoogleDrive

cd ./inference_code
python inference_seg.py -c config/seg_x8_superhuman.yaml # for 8x EMSR segmentation using superhuman model
python inference_seg.py -c config/seg_x8_MALA.yaml # for 8x EMSR segmentation using MALA model
python inference_seg.py -c config/seg_x16_superhuman.yaml # for 16x EMSR segmentation using superhuman model
python inference_seg.py -c config/seg_x16_MALA.yaml # for 16x EMSR segmentation using MALA model

AC4

Superhuman	MALA
GoogleDrive	GoogleDrive

Pre-trained models

We provide pre-trained models for all three stages.
We also provide pre-trained VGG model and SPyNet model.

CREMI	AC4
GoogleDrive	GoogleDrive

16x EMSR

8x EMSR

Contact

If you have any problem with the released code, please contact me by email (shoujt@mail.ustc.edu.cn).