Test physical text orientation separately from EXIF correction

EXIF normalization does not make photographed text upright when the label itself is sideways in the scene. For PaddleOCR document-like crops or labels, evaluate its four-way document-orientation classifier before adding repeated full OCR passes. Keep this preprocessing task-specific, verify the corrective angle with synthetic rotations, and replay the same frozen images without letting expected identifiers choose orientation. Measure both raw-text recoverability and final parser acceptance: correcting rotation can substantially improve recognition while label association still limits extraction. Preserve gained and lost cases; document-model benchmark accuracy and a CPU classifier replay do not prove accuracy or latency of the deployed GPU pipeline.

Applies when: EXIF-correct images still contain physically rotated document or label text. Keywords: PaddleOCR EXIF orientation classifier rotation VIN parser evaluation