Deep Autofocus with SEM

This repository contains the code for the paper Robust Autofocus Score Prediction of SEM on Magnification Based on Deep Learning System presented at Microscopy & MicroAnalysis (M&M 2020).

Table of Contents

• Environment setting
• Training a model
• Validation
• Conventional algorithms
• Results
• Citation
• Reference

0. Environment setting

#gpu
CUDA                    10.0
cudnn                   7.6.0

#package
pytorch                 1.1.0
torchvision             0.3.0
numpy                   1.16.4
matplotlib              3.1.1
pillow                  6.1.1

1. Training a model

1-1) Datasets

• Datasets are prohibited to share.
• Datasets should be like this if you have datasets and want to run training code.

 mnm2019_dataset/
    ├── 01220_grid+tinball
    ├── grid_tinball_zero_20200206_refined
    ├── Data_0
    └── cache.txt

 mnm2019_dataset/
    ├── train/
    ├── test/
    └── cache.txt

#head of cache.txt
24501	2734	3037                # number of train, valid, test
Data_0/169.jpg	0	500	0   # path_of_image, experts_score, magnification, specimen
Data_0/362.jpg	0	500	0
Data_0/420.jpg	0	2000	0
Data_0/402.jpg	0	1000	0
...

1-2) How to train

python train.py <opt>   # opt: a, b, c, d

• a: proposed model with new dataset
• b: proposed model with old dataset
• c: resnet50 with new dataset
• d: resnet50 with old dataset

2. Validation

2-1) Download pretrained model

Please download below models and make folder pretrained/ and place models in that folder by unzipping mnm2020_models.zip.

• models : Google Drive

 pretrained
    ├── mnm2020_models.zip
    ├── mnm2020_with_new.pth 
    ├── mnm2020_with_old.pth 
    ├── mnm2019_with_new.pth 
    └── mnm2019_with_old.pth 

2-2) How to test

python test.py <opt>    # opt: a, b, c, d

• a: proposed model with new dataset
• b: proposed model with old dataset
• c: resnet50 with new dataset
• d: resnet50 with old dataset

If you want to test log,

python test.py <opt> > <name>.txt

3. Conventional algorithms

There are 6 algorithms of sharpness functions we implemented. You can find those in af/sharpness.py. To compare performance, we only use variance absolute which is the best performing conventional algorithm in previous study.

cd af/
python autofocus.py <dataset path> <sharpness function> 
#Example: python autofocus.py ../../data/dataset/

4. Results

You can see our saved results in results/.

#(option) filename.txt
(a) mnm2020_with_new.txt
(b) mnm2020_with_old.txt
(c) mnm2019_with_new.txt
(d) mnm2019_with_old.txt
(var_abs) var_abs.txt

5. Citation

@proceeding{deep-autofocus,
  title         = {Robust Autofocus Score Prediction of SEM on Magnification Based on Deep Learning System},
  author        = {Moohyun Oh, Jonggyu Jang, Hyun Jong Yang, Hyeonsu Lyu},
  journal       = {arXiv or google scalar, **NEED TO UPDATE LATER}
  year          = {2020},
  howpublished  = {\url{https://github.com/blacknwhite5/deep-autofocus}}
}

6. Reference

• pytorch/vision/resnet.py