I am working with NIST Special Database 4, which is a database of fingerprints. The objective is to train a convolutional network (CNN) and train it to classify the fingerprints.

After I looked through the database, I notices some images are lighter and some are darker, an example is provided below:

Darker Image

Lighter (bleached) image

Now my question is if the images are used with unequal intensity (without any preprocessing), would that affect the training process for the CNN and subsequently the accuracy of the classification process?

Would Gamma Correction of the images to uniform intensity be better before the training & classification process?

Is there some best practices for Gamma correcting multiple images to similar intensity.

Would really appreciate some help on this.

  • $\begingroup$ May I know the reason for down-voting $\endgroup$
    – Ironluca
    Apr 14, 2017 at 7:13

1 Answer 1


In the Tensorflow example for CNN a distorted_inputs function is implemented. Besides flipping the images randomly, it's also randomly distorting

  • Image brightness
  • Image contrast

So, the intensity of the images is changed intentionally. They do it to increase the number of training images artificially.

Because of that knowledge I wouldn't suggest making the input similar. Moreover, I think the variability within the input dataset is essential for making the generalization, which is needed for classifications of that example.

Due to low reputation I am not able to post this information as a comment. Please be gently.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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