Enhance neural net measure detector with image-processing based measure detector

[Archer6621]

Currently the neural net measure detector has proven to be imprecise. Internally we have been working on an image-processing based solution, which ultimately also seems to have unsatisfactory precision and recall. However, since the image-processing solution uses a white-box approach to the problem, we will be able to tweak it in predictable ways to perform as needed.

We've identified that both measure detectors detect some false positives, but they are different in nature for both of them. The image-processing based solution also misses some measures, but can be made more sensitive to detect them anyway, at the cost of precision. The neural net solution seems to have very high recall, as it hasn't missed a measure in scores that we have provided so far. Since the imprecise results of the two detectors do not seem to overlap, they can be ruled out. This should improve precision tremendously and give us much better results.