unet-sem

image segmentation for CVD graphene SEM images
Based on: https://github.com/zhixuhao/unet

Usage

main.py [-h] [--crop] [--shuffle_data] [--augment_after AUGMENT_AFTER]
           [--augment] [--max_crop] [--crop_size CROP_SIZE]
           [--input_size INPUT_SIZE] [--ngpu NGPU] [--nepochs NEPOCHS]
           [--batch_size BATCH_SIZE] [--split SPLIT] [--lr LR]
           [--input_dir INPUT_DIR]

optional arguments:
  -h, --help            show this help message and exit
  --crop                Constructs batch using random crops.
  --shuffle_data        Shuffles data paths to switch what is in test/train.
  --augment_after AUGMENT_AFTER
                        Start augmenting data ater specified epoch,
                        inclusively.
  --augment             Constructs batch using augmentations.
  --max_crop            Crops using the maximum square size for each image.
                        Crop size is ignored.
  --crop_size CROP_SIZE
                        Size of cropped sample.
  --input_size INPUT_SIZE
                        Model input size. Cropped images will be rescaled to
                        this size.
  --ngpu NGPU           Number of GPUs.
  --nepochs NEPOCHS     Number of epochs.
  --batch_size BATCH_SIZE
                        Number of samples per batch.
  --split SPLIT         If float, fraction of data to use for validation. If
                        integer, number of folds. If zero, train on all data
                        (used for final model.
  --lr LR               Learning rate.
  --input_dir INPUT_DIR
                        Directory to pull images from

Image Data

Images used to train the neural network are available in the data directory:

• All directories that end in _Bad contain low fidelity masks and their corresponding images
• All directories that end in _Good contain high fidelity masks and their corresponding images
• The Old_data directory also contains high fidelity masks and their corresponding images

Use the fileproc.py script to combine all the images into a single directory and rename them correctly

Domain Size Analysis

domain_size_analysis.ipynb is the jupyter notebook that contains the code to our method of domain size analysis. It has the following sections:

• Import packages: imports all relevant packages
• Define all analysis functions: All the required classes and functions for domain size analysis are defined here
• Analysis of randomly generated data: A data generator is created to test the technqiue on artificially generated data
• Follow process for one simulated image: Run through the analysis process step by step for only one image (might be broken)
• Analysis of true data: Analysis of true data. User must select a directory that contains all the segmented masks
• Experimental section: Code blocks used to trace out code in case of problems