If you want to get some help, please be MUCH more specific.
Attach a typical image showing your problem.
Be precise about the concepts you are using. For example, what do you mean by "severity score"?
If possible, add links to these concepts or papers describing what you want to accomplish (no matter if its an old method, at least it will explain the purpose in detail even if the technique is outdated).
from my experience it does not make sense to apply rule learners (like RIPPER from RapidMiner) to images straight away.
You should use pre-processing (expert knowledge tuned) to extract some features first. After feature extraction you will be able to train a model with rules of the kind:
IF feature f1 > threshold t2, THEN severity score = 0.9
Apart from that, nowadays for this kind of problems Convolutional Neural Networks and Deep Learning are regarded as most suitable.