Determine the best features of an object given several images

asked 2020-01-15 19:25:36 -0500

I am trying to automate a form scanner and align new images of the same printed form. I'm using the orb descriptors of the correctly aligned form and matching it (with hamming distance) against the new images, which works decently. But, how could I, given several images of the correct form, extract the most consistent and best features for alignment to use with new images?

Thank you!

edit retag flag offensive close merge delete


Well in general youi do it by picking the ones which leads to the best results. You could make a statistic analysis - there are several algorithms to do it.

How about this: You first collect the data - the features(think about how to transfer your features into numbers) and the result value they lead to (for example the accuracy).

You then either use linear regression(this will allow you to predict the accuracy for features) or classification(you define for example accuracy < 0.5 as not good and accuracy >= 0.5 as good(binary classification - also called logistic regression) and this way you can classify new features as usefull or not) to do the job.

holger gravatar imageholger ( 2020-01-16 07:46:03 -0500 )edit