Idea: Apply shear+zeropadding to an image

[kolibril13]

Description

It is already possible to shear images in skimage (example).
However, some information of the image will get lost during the shear, when the shape of the plot is preserved.
In some cases it might be useful to preserve all image information and therefore change the shape of the plot with some zero-padding.

Suggestion:

An option for padding zeros when shearing an image.

Implementation

One naive way to do so is like this:

from matplotlib import pyplot as plt
import numpy as np

from skimage import data
import cv2

camera = data.camera()
camera = cv2.cvtColor(camera, cv2.COLOR_RGB2RGBA)
plt.imshow(camera, interpolation='nearest', cmap='gray');

shear_factor= 1.5

pad= np.full((int(camera.shape[0]/shear_factor), camera.shape[1]) ,0)
new_pad = cv2.cvtColor(np.uint8(pad), cv2.COLOR_GRAY2BGRA)
new_pad[:,:,3]=0 # setting alpha channle to zero
camera_new=np.concatenate((camera ,new_pad) ,axis=0)
plt.imshow(camera_new, interpolation='nearest', cmap='gray');

camerax = []
for i,row in enumerate(np.rot90(camera_new,1)):
    camerax.append(np.roll(row,int(i/shear_factor),axis=0))
camerax=np.uint8(np.rot90(camerax,-1))
plt.imshow(camerax, interpolation='nearest', cmap='gray');

Further notes:

I found #1779 , #1975 and #3548 as issues and pr that also address the shear of images.

[jni]

Thanks for raising this @kolibril13!

The real issue is that we don't have any support at all for negative coordinates, so any shear that extends into negative territory will be lost. So the way scikit-image would see your output is as a shear + translate down by ~350 pixels. This is readily expressed by an affine transform in scipy.ndimage.

Long term, I think the solution is for affine_transform to take in not just output_shape, but rather output_coordinates.

Anyway, this is all tricky and will require some careful thought for the API, but we'll definitely work on improving things here!