OCR extraction from poor-quality rasterised PDFs for GIS attribution

In all honesty if it were me I’d ask Claude what it recommends (if you haven’t already) - if you can give it some sample data, let it know you want it in FME it should give you a pretty decent start. 

 

Georeferencing is going to be the hardest or most manual step if that’s all the information you have to go off, so first of all consider if there’s anything else you could mark or label at that point to help this out.

Note that line colour is not entirely reliable because in the pdf it is covered by labels, project boundaries etc. And note diameters are nominal, a mix of internal/external d. You might be able to infer PE100 pipe material from that?

After georeferencing the image, 1 metre square buffer/clip out the image around a mid-section snip of each of your mains pipes which are more than ~10m long (ie start snip at 1 metre, end at 1 metre before end). The image in each of these buffer/clips should be some part blue and to be confident it contains a pipe, count the blue pixels, or see if they generalise to a straight line inside the buffer. Then for the each clipped images tesseract it or pass it to an image model with ollama to get the text in it (maybe prompting for text to be oriented with a blue line). Image models might handle rotated and low-res text better.

Very much depends entirely on georeferencing accuracy, and somewhat that your GIS pipe geometries match to the PDF to begin with. And you’ll need to fill in the gaps along the mains for those pipes between valves without a label in the pdf.