To improve the processing of NUM2 ionograms

The default _keras_ocr_ text recognition model that was used could be further trained to better read NUM2 metadata numbers, and to only read numbers and not letters. For instance, there were at least 23,984 NUM2 images that seemed to adhere to the metadata format (i.e: had 15 characters beginning with ‘10’), but then had certain numbers misread as letters. There are also several instances where an image that is actually of NUM2 type was not read, perhaps due to scan quality issues. Further training of the text recognition model could improve the reading accuracy.

Start by locating the entry point that uses the default keras_ocr text-recognition model and follow how NUM2 metadata is identified from scanned ionogram images. Review the examples of 15-character values beginning with “10,” then define evaluation data for misread letters and missed NUM2 images; done means improved recognition accuracy without treating letters as valid numbers.

Written by the indexing model from the issue text.