Bo Andersen

I test extraction against the failure cases first, because an OCR engine that reads a clean export is easy and one that reads a photograph taken in poor light is the reason anybody needs it. Order of operations matters more than it looks. Recognising text in the original language and then translating produces a different result from translating a page and recognising the result, because the second step has errors to work with and the first does not. Where the source language is known I extract first. Image quality sets the ceiling and there is no recovery above it. Rotation applied a fraction of a degree off, uneven lighting, a low contrast between ink and paper and compression artefacts all reduce accuracy, and each is cheaper to fix at the source than to correct afterwards. Screenshots have their own difficulty. Sub-pixel antialiasing produces letterforms the engine has never seen, and a screenshot scaled to a fraction of its original size is worse still. Capturing text at its native size matters more than it appears. Handwriting has a much lower ceiling and no general solution. I am direct about which kinds of handwriting work and which do not, and I do not present a result as reliable when it is a guess. Every page reports the confidence and shows what came out wrong, because a transcription presented as exact and containing one wrong digit is worse than one flagged as uncertain. Where the text is a table, the recognition is the easy half and the reconstruction is the rest. A flat stream of values with the column boundaries lost is not a table, and any tool that presents it as one is doing layout analysis with a low success rate that the output does not advertise.