Tensorflow, Keras: Tensor normalization by axis

Question

Assume we have images Tensor A with shape (None, 200, 200, 1) . where None is the batch size, and (200, 200, 1) is the image size.

How to perform normalization (0 to 1) on each image (not using for iteration)?

That is:

A[0] = (A[0] - A[0].min()) / (A[0].max() - A[0].min())
A[1] = (A[1] - A[1].min()) / (A[1].max() - A[1].min())
...
A[n] = (A[n] - A[n].min()) / (A[n].max() - A[n].min())

if I just directly use A = (A - A.min()) / (A.max() - A.min()) , it will normalize all the images by the global max and min . I wish to normalize each image with their own max and min .

In other words, how to achieve a max or min operation, that results in a shape of: (None, 1, 1, 1) , where each (1, 1, 1) contains the max or min of each image.

Answer 1

You can use tf.reduce_max and tf.reduce_min .

import tensorflow as tf

A = tf.random_normal(shape=(-1, 200, 200, 1))
B = tf.reduce_max(A, axis=(1, 2, 3))
C = tf.reduce_min(A, axis=(1, 2, 3))

print(B.shape)
print(C.shape)

Output:

(?,)
(?,)

Also, the output in your case needs to be of the shape (None, 1, 1, 1) , not (None, 1, 1) , as you have included the last channel dimension.

B = tf.reshape(B, (-1, 1, 1, 1))
C = tf.reshape(C, (-1, 1, 1, 1))

print(B.shape)
print(C.shape)

The above code gives the following output:

(?, 1, 1, 1)
(?, 1, 1, 1)

Finally, why are you subtracting the max from each pixel, do you mean to subtract the min for normalizing each image between 0-1 ?

D = (A - C) / (B - C)
print(D.shape)

gives

(?, 200, 200, 1)

as expected.