@Frank is right, the training data at hand is not a good use case for Minos. Minos, after all, like ShapeFinder and many other pattern recognition tools, tries to build a feature-based description of the object(s) the classifier is supposed to be able to identify. For Minos, the basic features from which a model is constructed are contrast-based geometrical descriptions of dark/bright regions in an object. This means, that in order to build a suitable feature description for a given model - suitable in this context meaning that the feature description matches each and every instance that has been trained for this model - it’ll be necessary to find features that match every instance of a model.
This task, as you can imagine, is made horribly difficult by geometric variations (which is incidentally the reason why Minos is not without some additional tricks capable of working scale and/or rotation invariant) and even more so by an unpredictable constellation of features. The latter is usually the case when looking at characters generated by cheap industrial ink jet printers like the ones in your training set.
Consider for example the following specimens of the model for 6:
All these images have been added to the same model - yet finding features that are common to all of these four images is nearly impossible (and effectively what Minos will find on these is descriptive for noise and background rather than the object you’re looking at in this case). This alone might already give rise to the behavior you’re seeing (the defined, albeit somewhat poorly described interruption of the training process).
To further complicate things, your training set contains a fairly large number of models (most of which exhibit to some extent the problem pointed out on model 6), which adds to the problem in that it’ll now not only be difficult to find feature sets common to all instances of a given model, but also features that will not introduce ambiguities between different models.
The two factors together - the poorly reproducible feature configuration inside the trained models plus the large number of models to be distinguished - bring up the behavior you are seeing: The interruption of the learning process. Unfortunately, the underlying function
LearnCLFFromMTS only returns a boolean value, which makes it hard for the TeachBench to give further diagnostic information. On the other hand, luckily there are only two cases I know of when the classifier generation fails:
- If Minos tries to reach outside the images for feature extraction (this can happen when making too generous use of the invariants generation feature while having at least one instance lying close to the edge of the image - therefore @Frank’s questions…).
- In a situation where the generation of a suitable and unambiguous set of features is not possible.
Luckily both of them don’t occur too often
The bad news for you is that it’s - at least to my opinion - not possible to successfully generate a worthwhile classifier for your use case with Minos. I have made good experience with Manto on cases like this, and its successor Polimago will most likely perform equally well if not better (but it’ll also require a significant number of training images (but the 948 instances that you have already trained are no trifle either…). Should you already have purchased Minos at this point then please contact your sales channel so that we can work out a solution for you.