Distorting the MNIST Image Data Set

I wrote an article titled “Distorting the MNIST Image Data Set” in the July 2014 issue of MSDN Magazine. See http://msdn.microsoft.com/en-us/magazine/dn754573.aspx.


The NMIST (mixed National Institute and Standards and Technology) image data set is a collection of small (28 by 28 pixels) black-and-white images of handwritten 0-9 digits. There are 60,000 images in a training set and another 10,000 images in a test set.

The data sets were created so that image recognition researchers and practitioners can test new algorithms against the set, and compare their results with other results.

Even 60,000 training images aren’t very many, so a clever idea is to create additional training images from the original images, by slightly distorting each image. For example, the screenshot above shows the third training image, which happens to be a ‘4’, and a distorted version.

Distorting images is surprisingly tricky. There are many possibilities, but the technique I demonstrate uses a Gaussian kernel, which is an interesting topic in its own right.

This entry was posted in Machine Learning. Bookmark the permalink.