crop_image.py

from skimage.measure import regionprops, label

def crop_image(x, threshold=230, plot=True):
  ''' 
  finds large connected components using skimage.measure.label,
  and displays them with matplotlib.pyplot.
  '''
  im = train_images[x]
  mask = np.where(im > threshold, 1, 0)
  # now get label_im, a version of the thresholded image, but
  # where insted of being 1 or 0, each connected component 
  # has a different integer value "label"
  label_im = label(mask, connectivity=2) # could also be 1-connectivity...
  # this is useful: skimage.measure.regionprops will let us measure area
  regions = regionprops(label_im)

  if plot == True:
    print('Image number',x,'... True answer',train_y[x])
    fig, axs = plt.subplots(1,2)
    axs[0].imshow(im,cmap=plt.cm.gray_r)
    axs[1].imshow(label_im,cmap=plt.cm.magma_r)
    plt.show()

  for region in regions:
    if region['Area'] < 50:
      continue
    miny, minx, maxy, maxx = region['BoundingBox']
    # get centre
    cx, cy = int((minx+maxx)/2), int((miny+maxy)/2)
    plt.figure(figsize=(1, 1))
    # if bounding box will go outside image, then shift
    if cx < 14:
      shift = 14 - cx
      cx += shift
    if cy < 14:
      shift = 14 - cy
      cy += shift
    # show the found numerals (just by cropping label_im)
    if plot == True:
      plt.imshow(label_im[cy-14:cy+14, cx-14:cx+14],cmap=plt.cm.gray_r)
      plt.show()