Standardized Spectrogram Generation for Image Classification with CNN

Question

I am generating spectrograms for an image classification task using CNN-based VGG16.

I have data from multiple individuals. I am basically trying to detect an activity pattern in EEG signals - these activity patterns appear in a certain frequency range.

Depending on the strength of the activity, the pattern is either yellow-ish (using cmap as colormap) or red. I am facing the following obstacles: (Just for information, in cmap, red is used to display strong PSD values, conversely the yellow)

• Patterns might differ depending on the individual, external reasons, channel used, etc. Therefore, it can sometimes be around 15Hz or 30Hz, sometimes very strong with red color (high PSD), sometimes low, as any other brain activity.
• Since I have artifacts (high amplitude), sometimes these become more dominant on the real signal and choosing the right color map becomes even harder.

The problem arises when I generate spectrograms. Because of the vmin, vmax value I am using (the same value for every case), I am unable to distinguish some patterns - that are displayed as yellow, rather than red.

My question is, how can I set the vmin and vmax value so that the patterns become visible in any case? Or, would CNN focus more on the "pattern" itself that the colormap becomes less important than I think?

Thank you in advance for any kind of contribution.