E N D
I’ve been kicking around a coding project that is a new method for object recognition. There is a classic data set for object recognition; a set of thousands of individual handwritten digits scanned from postal zip codes. My goal is to create a program that can correctly identify individual digits. My method takes a test digit and rotates it through a couple of angles, each time generating a univariate PDF. I can get moderately good results by looking at the sum of the square of the residuals of QQ plots between archetypes for each digit and the test digit, but I’d like to see how I can use Bayesian methods to improve my accuracy. N.B. as with all things, there already exist very effective techniques for digit recognition using neural networks and related tools. My coding project is just trying out a new idea. Example of some of the digits in the database Comparison of the PDF (red) for a randomly chosen test digit ‘3’ and the archetype PDF for ‘3’ (blue) Adding small amounts of shear and/or rotation to a test digit • My question: How could Bayesian methods fit into this? • Estimate probability a new object is a certain digit (how is this Bayesian?) • Estimate optimal joint rotation and shear given a guess of the digit? • Probabilistic network between observations (PDFs) and IDs (digits) (what does this even mean?, is it Bayesian?)