OCR for Scanned Documents

The contrast this whole course has been building to

scanned_notice.pdf isn't a native-text PDF at all: it was built by rendering a PNG image (sample_data/_src/scanned_notice.png) into a PDF page. There is no PDF text layer in it, no TextItem LiteParse could ever find by reading PDF text objects, because none exist. To a computer, without OCR, it's just a picture. This is the exact situation ocr_enabled (mentioned since Lesson 2) exists for.

ocr_enabled=False: nothing to extract

With OCR off, LiteParse only reads native PDF text objects. This page has zero of them, so the result is an empty string, correctly: there really is no extractable native text on this page, no matter how obviously readable the notice is to a human looking at it.

ocr_enabled=True: renders the page, runs Tesseract

With OCR on (the default), LiteParse's complexity heuristics recognize this page needs more than the native-text path: no native text and a full-page embedded image is exactly the signature of a scan. It renders the page to a bitmap and runs Tesseract OCR against that bitmap, recovering real, readable text: "NOTICE... Building maintenance will occur on Saturday... Water service will be interrupted from 9 AM to 1 PM..." The OCR text isn't pixel-perfect (spacing and line breaks don't exactly match the source image), but the words are all there and correct.

is_complex(): seeing the decision, not just the result

stats = parser.is_complex(SCANNED_PDF)

is_complex() runs the same cheap pre-check LiteParse's own OCR decision is based on, without doing a full parse. On this page it reports needs_ocr=True with reasons ['no-text', 'embedded-images']: zero native text characters, and a page that's essentially a picture. This is useful on its own, as a fast way to triage a folder of PDFs into "needs OCR" and "doesn't" before committing to a full (slower) OCR parse of everything. Lesson 11's checkpoint project builds exactly that triage step, though it also shows a case where needs_ocr fires on a page that already has perfectly good native text, so it's a signal to weigh, not a rule to follow blindly.