ID3 classifier to classify MNIST handwritten digits dataset

Question

I'm implementing the ID3 (Decision Tree) algorithm for multiple labels, to classify MNIST handwritten digits dataset which contains 28*28 pixels with values 0-255 where 0 represents background and 255 represents foreground.

I'm trying to find a set of attributes that will get me a low error rate. Currently, i'm using threshold of 0 for each pixel and I get an error rate of 11%.

I would like suggestions or ideas to improve the error rate by setting a new set of attributes (i was thinking about detecting curves and lines in the image, however I cannot seem to find the way to do so in JAVA).

Thanks.

Answer 1

We have found out that dividing the image to frames (between 4 to 8) helped improving the predict %. Also, we have added features such as lines, curves and such.