Single-frame super-resolution microscopy（SFSRM）

Example data and demo of SFSRM for single-frame super resolution of microscopy images.

Requirements

SFSRM is built with Python and pytorch.

• Python >= 3.7
• PyTorch >= 1.7

  Install

  1. Clone repo
     

     git clone https://github.com/crrayna/SFSRM.git

  2. Install dependent packages
     

     pip install -r requirements.txt

     Quick inference

  3. Download pretrianed models
     

Download pretrained models at https://drive.google.com/drive/folders/1UnaDwrt1FNSAUT_OlosqvoV4jsxIxIhi?usp=sharing and put them in the pretrained_network folder

2. Inference
   

   python test.py -opt options/test/test_example_microtubule.yml

   Results are in the results folder

*For your own test data, we recommend using the SRRF plugin in FIJI/ImageJ to generate the edge map. The plugin provides a 32-bit SRRF image. You will need to convert this image to an 8-bit edge map. Prior to the conversion, it may be necessary to adjust the dynamic range to ensure that the background intensity of your edge map matches the level of our sample edge map. Please note that the background intensity can vary for different samples (e.g., MT, mito, ER), so adjustments might be needed accordingly.

Acknowledgements

The codes are based on ESRGAN and unetgan. Please also follow their licenses. Thanks for their awesome works.

References

[1] Wang, Xintao, et al. "Esrgan: Enhanced super-resolution generative adversarial networks." Proceedings of the European conference on computer vision (ECCV) workshops. 2018.
[2] Schonfeld, Edgar, Bernt Schiele, and Anna Khoreva. "A u-net based discriminator for generative adversarial networks." Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2020.

Contact

If you have any questions, please email meshyao@ust.hk