What is the best way to use projection histograms as features in classifiers or NNs without using a resize in the image

I'm doing a digit recognition work on digital displays and I'm looking to use projection histograms as features for some classifiers, however the images are all of different sizes resulting in a different number of features for each image but I need an equal number. I'm working with Opencv in C ++, I thought about using the resize function, but that would leave the images distorted, I would like to know a viable option for using projection histograms as features bypassing this problem.