Tesseract default --psm 3 misses sparse callout/label text on large engineering plan sheets
Tesseract with default --psm 3 misses sparse callout/label text on large engineering plan sheets (e.g. 10820x7236 px, 300 dpi construction drawings): image_to_data returned only ~545 words and did not detect a clearly visible road-name label, so term-location for redaction QA failed with 'no OCR box match'.
Run a second pass with --psm 11 (sparse text) and union the word boxes with the default pass. On the same sheet, psm 11 found 1409 words vs 545 and located the missed label exactly ('Neabsco' at (2271,3120,185,42)). Page segmentation, not resolution, was the limiter: default psm 3 assumes document layout and discards isolated callouts on drawings that are mostly linework. Practical recipe for locating known terms on plan sheets: words = psm11_words + psm3_words; match single tokens by casefolded substring, multi-token phrases by fuzzy window (rapidfuzz ratio >= 82) including a collapsed no-space variant, since tesseract splits/joins tokens unpredictably on drawings. Also lift PIL's decompression-bomb cap (Image.MAX_IMAGE_PIXELS = None) for full-sheet renders.