By the definition of min-max normalisation, the value is divided by max - min, what if the max and min have the same value?

So I have read about this anwser, but I am still not sure about my case.

For example, I have an image with n channels and I want to perform min-max normalisation over each channel i.e. $$x_{channel_a} = \frac{x_{channel_a} - min_{channel_a}}{max_{channel_a} - min_{channel_a}}$$

It is possible that only channel a is the same and other channels are different. Does it mean I should drop this channel information? How exactly can I do? What about setting the channel value to 0 if all channel values are the same?

If I drop this value, will it cause any loss of information. For example, I have a image with pixels value in RGB format [12, 23, 34], [12, 25, 87], [12, 182, 230]. Since R channel is the same, I can just drop it and the image become to [0, 23, 34], [0, 25, 87], [0, 182, 230]. Is this correct?

By the way, I am trying to use ResNet-18 to extract the image features.

  • 2
    $\begingroup$ 0/0 is undefined $\endgroup$ – Glen_b Dec 18 '19 at 8:13
  • 1
    $\begingroup$ When you are willing to standardize your variables and max==min, you have no information. Equivalently, you may replace the common value by any constant number you wish. $\endgroup$ – whuber Dec 18 '19 at 14:47
  • $\begingroup$ @Glen_b-ReinstateMonica That is the problem. I am trying to verify is that ok to set the value to any valid constant I want so that I can avoid dividedByZero error. $\endgroup$ – Dylan Wang Dec 19 '19 at 0:28
  • $\begingroup$ @whuber Thank you for clarifying that. $\endgroup$ – Dylan Wang Dec 19 '19 at 0:28
  • $\begingroup$ You could make a case for it being any number; 1/2 would be an obvious value to use in some situations, but any other value could be just as meaningful / equally meaningless in another situation. I'd be inclined to just use "missing" (however that's handled in whatever you're using) for those values. $\endgroup$ – Glen_b Dec 19 '19 at 1:59

If $\min_i x_i = \max_i x_i$ then this implies that $x_1 = \cdots = x_n$ (i.e., all data values are the same). In that case, there is not really any difference to "normalise". By convention, you would probability set the "normalised" values to $z_1 = \cdots = z_n = 0$ in this case, which shifts them to a mean of zero, and retains the property that:

$$z_k (\max_i x_i - \min_i x_i) = x_k - \min_i x_i.$$

| cite | improve this answer | |

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.