Elastic transformation of an image in Python

elastic_transform.py

import numpy as np
from scipy.ndimage.interpolation import map_coordinates
from scipy.ndimage.filters import gaussian_filter

def elastic_transform(image, alpha, sigma, random_state=None):
    """Elastic deformation of images as described in [Simard2003]_.
    .. [Simard2003] Simard, Steinkraus and Platt, "Best Practices for
       Convolutional Neural Networks applied to Visual Document Analysis", in
       Proc. of the International Conference on Document Analysis and
       Recognition, 2003.
    """
    if random_state is None:
        random_state = np.random.RandomState(None)

    shape = image.shape
    dx = gaussian_filter((random_state.rand(*shape) * 2 - 1), sigma, mode="constant", cval=0) * alpha
    dy = gaussian_filter((random_state.rand(*shape) * 2 - 1), sigma, mode="constant", cval=0) * alpha
    dz = np.zeros_like(dx)

    x, y, z = np.meshgrid(np.arange(shape[0]), np.arange(shape[1]), np.arange(shape[2]))
    print x.shape
    indices = np.reshape(y+dy, (-1, 1)), np.reshape(x+dx, (-1, 1)), np.reshape(z, (-1, 1))

    distored_image = map_coordinates(image, indices, order=1, mode='reflect')
    return distored_image.reshape(image.shape)

hi!
How do you apply the same transformation to the mask ? I have a set of satellite images with each their corresponding road segmentation (black and white, (400,400))

Same question here, the implementation is inherently randomized for each run, making it impossible to apply the exactly same transform to both the original image and the mask image.

Just fix the random_state for both calls. :)

Can I apply it for a multi-class dataset for the segmentation task?

Can you elaborate a bit more @jepperaskdk - i would love to make use of this however im unsure what you mean in terms of stack the two images (i..e pull in the original from one folder X_img and the corresponding mask Y-Img and separate after processing? Literally run them all through and then manually / script separate the images that are the output?

Can you elaborate a bit more @jepperaskdk - i would love to make use of this however im unsure what you mean in terms of stack the two images (i..e pull in the original from one folder X_img and the corresponding mask Y-Img and separate after processing? Literally run them all through and then manually / script separate the images that are the output?

On second thought, I'm not sure if it works.

Hi, I have load a RGB img whose shape is (400, 248, 3), but I have got an error
ValueError: operands could not be broadcast together with shapes (248,400,3) (400,248,3)
in the code here
indices = np.reshape(y + dy, (-1, 1)), np.reshape(x + dx, (-1, 1)), np.reshape(z, (-1, 1))
can anyone help ? THX!!!

you need to invert the shapes in the resolution of x,y,z :
x, y, z = np.meshgrid(np.arange(shape[1]), np.arange(shape[0]), np.arange(shape[2]))
instead of
x, y, z = np.meshgrid(np.arange(shape[0]), np.arange(shape[1]), np.arange(shape[2]))

Inverting the shapes flipped the image for me. Setting the the indexing of the meshgrid to 'ij' instead fixed this issue:
x, y, z = np.meshgrid(np.arange(shape[0]), np.arange(shape[1]), np.arange(shape[2]), indexing='ij')

Since the interpolation is also done over the channel dimension, I got this interpolation artifacts.

So I decided to do the interpolation channel-wise.
https://gist.github.com/mvoelk/0880f5de7c101c093165e1e46ce3f6e5

which augement did you apply for dataset @bigfred76