CS-Sim: synthetic images with curvilinear structures

Package for generating synthetic images with curvilinear structures, and corrupting them with noise, background, and blur.

The images are intended to be used as ground truth for filament- and curvilinear-structure-detection methods.

Installation

pip install git+https://github.com/amedyukhina/CS-Sim.git

Usage

See the demo notebook for details.

Generate synthetic image

from cs_sim.synth.filaments import generate_img_with_filaments

img = generate_img_with_filaments(imgshape=(20, 100, 100), n_filaments=10, maxval=255)

Corrupt image with noise, blur and background

Specify corruption steps

There are 4 steps that can be combined in any order:

Perlin noise: adds low-frequency background. The size parameter specifies the size of the low-frequency pattern (in pixels), the value parameter specifies the amplitude.
Convolve: convolves the image with either a specified PSF image (psf parameter), or with a gaussian kernel of specified sigma.
Poisson noise, with specified snr.
Gaussian noise, with specified snr.


corruption_steps = [
    ('perlin_noise', {'size': 50, 'value': 0.1}),
    ('poisson_noise', {'snr': 2}),
    ('convolve', {'sigma': 2}),
    ('gaussian_noise', {'snr': 100})
]

Corrupt the image

from cs_sim.corrupt import corrupt_image

img_corr = corrupt_image(img, corruption_steps)

Save parameters to a json file for batch processing

import json

params_synth_data = dict(
    imgshape=(20, 100, 100),
    n_filaments=10,
    maxval=255,
    nval=100
)

with open('parameters_synth_data.json', 'w') as f:
    json.dump(params_synth_data, f, indent=4)

corruption_steps = [
    ('perlin_noise', {'size': 50, 'value': 0.1}),
    ('poisson_noise', {'snr': 2}),
    ('convolve', {'sigma': 2}),
    ('gaussian_noise', {'snr': 100})
]

with open('corruption_steps.json', 'w') as f:
    json.dump(corruption_steps, f, indent=4)

Batch processing

Generating synthetic images

from cs_sim.batch.batch_synth import batch_generate_img_with_filaments

with open('parameters_synth_data.json') as f:
    params = json.load(f)
batch_generate_img_with_filaments(n_img=10, n_jobs=10, dir_out='test_input', **params)

Corrupting the images

from cs_sim.batch.batch_corrupt import batch_corrupt_image

with open('corruption_steps.json') as f:
    corr_steps = json.load(f)
batch_corrupt_image('test_input', 'test_output', corr_steps, n_jobs=10)

Using the scripts

python scripts/batch_generate_img_with_filaments.py -p parameters_synth_data.json -o test_input -n 10

python scripts/batch_corrupt_images.py -p corruption_steps.json -i test_input -o test_output