如何利用已知变换参数还原Keras中ImageDataGenerator的图像变换？

Absolutely! You can absolutely reverse the transformations applied by ImageDataGenerator—as long as you have the exact parameter values used for each specific transformation on a given image. Let’s walk through how to handle the most common operations, step by step:

1. Horizontal/Vertical Flips #

• If you know an image was horizontally flipped (from a recorded horizontal_flip event), flipping it again will bring it back to the original. Same logic applies for vertical flips—two flips cancel each other out.
• Quick code example:

  # Reverse a horizontal flip
  reversed_image = np.fliplr(transformed_image)
  
  # Reverse a vertical flip
  reversed_image = np.flipud(transformed_image)
  

2. Rotations #

• ImageDataGenerator applies clockwise rotations by default. To reverse a rotation of theta degrees, rotate the transformed image counterclockwise by the same theta (or clockwise by 360 - theta degrees).
• Critical note: Use the exact same fill_mode (like 'nearest', 'reflect', or 'constant') that was used in the original transformation—mismatching fill modes will lead to inconsistent edge pixels that can’t be perfectly reversed.
• Example code:

  from keras.preprocessing.image import apply_affine_transform
  
  # Original rotation was 30 degrees clockwise, fill mode 'nearest'
  reversed_image = apply_affine_transform(
      transformed_image,
      theta=-30,  # Negative value = counterclockwise rotation
      fill_mode='nearest'
  )
  

3. Width/Height Shifts #

• If an image was shifted horizontally by shift_x (a fraction of the image width) and vertically by shift_y (a fraction of the height), reverse it by shifting with -shift_x and -shift_y respectively.
• Again, stick to the original fill mode for best results:

  # Original shift: right 20% of width (shift_x=0.2), up 10% of height (shift_y=-0.1)
  img_width, img_height = transformed_image.shape[1], transformed_image.shape[0]
  reversed_image = apply_affine_transform(
      transformed_image,
      tx=-0.2 * img_width,  # Negative tx shifts left
      ty=0.1 * img_height,  # Positive ty shifts down
      fill_mode='nearest'
  )
  

4. Zoom #

• For a zoom factor z (where z > 1 means zooming in, z < 1 means zooming out), reverse it by zooming with 1/z.
• Make sure to center the zoom the same way ImageDataGenerator does (default is center-aligned) to keep the image properly aligned:

  # Original zoom was 1.5 (zoomed in 50%)
  reversed_image = apply_affine_transform(
      transformed_image,
      zx=1/1.5,
      zy=1/1.5,
      fill_mode='nearest'
  )
  

5. Shear Transformations #

• A clockwise shear of shear degrees can be reversed by applying a counterclockwise shear of the same magnitude (i.e., -shear degrees):

  # Original shear was 10 degrees clockwise
  reversed_image = apply_affine_transform(
      transformed_image,
      shear=-10,
      fill_mode='nearest'
  )
  

Important Things to Keep in Mind #

• Randomness is a dealbreaker without logs: If you used random transformations (like random rotation ranges, random flips), you must have recorded the exact parameter value used for each individual image. You can’t reverse a random transform if you don’t know which specific value was applied to that image.
• Fill mode consistency is non-negotiable: Using a different fill mode than the original transformation will create edge pixels that don’t match the original image, making a perfect reversal impossible.
• Don’t forget pixel scaling: If you applied normalization (e.g., rescale=1/255), reverse that first by multiplying the image by 255 (or the inverse of your rescale factor) before handling spatial transformations.

内容的提问来源于stack exchange，提问作者user782400