Labeling-Criteria.md

Labeling of Digital Images

For the recognition algorithm it is important to get a clear reading and also to get the information, if the number is not totally clear at the moment of reading. Therefore I have decided to extend the 10 digits by one additional category - the so called NaN (Not a Number). As there are all kind of digit types and reading conditions, I needed to define an easy criteria for distinguish them. I have decided to use the following definition:

Good images:

• Only one number is visible
• Number is totally visible
• There is at least one pixel distance to the boundary of the image

Examples:

​

Bad image:

• Digit only partial visible
• Digit is toughing the boundary of the image

Examples:

​

Non Unique Images - not trained (!)

There are some "border" images, which are not totally clear. I have decided to exclude them for the moment from the training. Never the less, these images are collected for later usage.

Examples:

​