Scipy rotate and zoom an image without changing its dimensions

Question

For my neural network I want to augment my training data by adding small random rotations and zooms to my images. The issue I am having is that scipy is changing the size of my

Answer 1

scipy.ndimage.rotate accepts a reshape= parameter:

reshape : bool, optional

If reshape is true, the output shape is adapted so that the input array is contained completely in the output. Default is True.

So to "clip" the edges you can simply call scipy.ndimage.rotate(img, ..., reshape=False).

from scipy.ndimage import rotate
from scipy.misc import face
from matplotlib import pyplot as plt

img = face()
rot = rotate(img, 30, reshape=False)

fig, ax = plt.subplots(1, 2)
ax[0].imshow(img)
ax[1].imshow(rot)

Things are more complicated for scipy.ndimage.zoom.

A naive method would be to zoom the entire input array, then use slice indexing and/or zero-padding to make the output the same size as your input. However, in cases where you're increasing the size of the image it's wasteful to interpolate pixels that are only going to get clipped off at the edges anyway.

Instead you could index only the part of the input that will fall within the bounds of the output array before you apply zoom:

import numpy as np
from scipy.ndimage import zoom


def clipped_zoom(img, zoom_factor, **kwargs):

    h, w = img.shape[:2]

    # For multichannel images we don't want to apply the zoom factor to the RGB
    # dimension, so instead we create a tuple of zoom factors, one per array
    # dimension, with 1's for any trailing dimensions after the width and height.
    zoom_tuple = (zoom_factor,) * 2 + (1,) * (img.ndim - 2)

    # Zooming out
    if zoom_factor < 1:

        # Bounding box of the zoomed-out image within the output array
        zh = int(np.round(h * zoom_factor))
        zw = int(np.round(w * zoom_factor))
        top = (h - zh) // 2
        left = (w - zw) // 2

        # Zero-padding
        out = np.zeros_like(img)
        out[top:top+zh, left:left+zw] = zoom(img, zoom_tuple, **kwargs)

    # Zooming in
    elif zoom_factor > 1:

        # Bounding box of the zoomed-in region within the input array
        zh = int(np.round(h / zoom_factor))
        zw = int(np.round(w / zoom_factor))
        top = (h - zh) // 2
        left = (w - zw) // 2

        out = zoom(img[top:top+zh, left:left+zw], zoom_tuple, **kwargs)

        # `out` might still be slightly larger than `img` due to rounding, so
        # trim off any extra pixels at the edges
        trim_top = ((out.shape[0] - h) // 2)
        trim_left = ((out.shape[1] - w) // 2)
        out = out[trim_top:trim_top+h, trim_left:trim_left+w]

    # If zoom_factor == 1, just return the input array
    else:
        out = img
    return out

For example:

zm1 = clipped_zoom(img, 0.5)
zm2 = clipped_zoom(img, 1.5)

fig, ax = plt.subplots(1, 3)
ax[0].imshow(img)
ax[1].imshow(zm1)
ax[2].imshow(zm2)

Answer 2

I recommend using cv2.resize because it is way faster than scipy.ndimage.zoom, probably due to support for simpler interpolation methods.

For a 480x640 image :

• cv2.resize takes ~2 ms
• scipy.ndimage.zoom takes ~500 ms
• scipy.ndimage.zoom(...,order=0) takes ~175ms

If you are doing the data augmentation on the fly, this amount of speedup is invaluable because it means more experiments in less time.

Here is a version of clipped_zoom using cv2.resize

def cv2_clipped_zoom(img, zoom_factor):
    """
    Center zoom in/out of the given image and returning an enlarged/shrinked view of 
    the image without changing dimensions
    Args:
        img : Image array
        zoom_factor : amount of zoom as a ratio (0 to Inf)
    """
    height, width = img.shape[:2] # It's also the final desired shape
    new_height, new_width = int(height * zoom_factor), int(width * zoom_factor)

    ### Crop only the part that will remain in the result (more efficient)
    # Centered bbox of the final desired size in resized (larger/smaller) image coordinates
    y1, x1 = max(0, new_height - height) // 2, max(0, new_width - width) // 2
    y2, x2 = y1 + height, x1 + width
    bbox = np.array([y1,x1,y2,x2])
    # Map back to original image coordinates
    bbox = (bbox / zoom_factor).astype(np.int)
    y1, x1, y2, x2 = bbox
    cropped_img = img[y1:y2, x1:x2]

    # Handle padding when downscaling
    resize_height, resize_width = min(new_height, height), min(new_width, width)
    pad_height1, pad_width1 = (height - resize_height) // 2, (width - resize_width) //2
    pad_height2, pad_width2 = (height - resize_height) - pad_height1, (width - resize_width) - pad_width1
    pad_spec = [(pad_height1, pad_height2), (pad_width1, pad_width2)] + [(0,0)] * (img.ndim - 2)

    result = cv2.resize(cropped_img, (resize_width, resize_height))
    result = np.pad(result, pad_spec, mode='constant')
    assert result.shape[0] == height and result.shape[1] == width
    return result