Denoising training images in Capstone Project

I’m trying to use non local mean(nlm) from skimage for denoising the training dataset and then I’m cropping the images to get the parts containing the text but it’s becoming too computationally expensive and I can’t process even a sixth of the images in 3 plus hours, I’m also running out of disc space in colab.
Is there a better way to include denoising in the pipeline of training?
I’m using albumentations for the transformations on the images.