I used images as labels for classifications instead of binary or integer classifications. I worked. But not so good as I had hoped.
Still I don’t give up. So what about x,y coordinates as labels. I was aiming at the textbook ?typical accuracy maps for construction.
I thought maybe every point could be drawn on a common image.
This way. A brain or my primitive model could adapt a little its predictions. Make some comparisons with earlier predictions.
The color for each prediction should be different. The labels are the black points. The numbers should not be visible on the image. Just the classification point in the corresponding color.
So with this my model will have continous update of the previous predictions via the classification image.
The result gave me an idea. Using KMeans for separation of the numbers there was an obvious class functionality missing in KMeans. The Surrounding class. Which is some kind of uncertainty. So we should have 11 classes 10 for the numbers and 1 for the surrounding class.