Intensity classification normalisation

Hi everyone,
I am facing a quite simple problem, yet I cannot solve it. I am trying to classify images according to their intensity. Min max normalization or division by max don’t work, as it “kills” the intensity differences. if I don’t normalize at all, it works like charm. The problem is that if I try to process data with lower intensity levels, hence a lower contrast, it only succeeds in classifying the brightest ones.
I am sure there is a way to normalize in a clever way so I do not have to include those new images to my training dataset every time I wanna apply my network to a new dataset. Any idea?
Than you!