UVDoc

UVDoc is an image-to-text model that performs document image unwarping — correcting geometric distortions such as curling, tilting, and perspective deformation in photographed documents to improve downstream OCR accuracy. It is part of the PaddleOCR toolkit and is hosted by gigarouter as a managed, OpenAI-compatible API. The model is designed to preprocess document images before text recognition, addressing common real-world issues where a document is photographed at an angle or on a curved surface. By applying a learned geometric transformation, UVDoc produces a flat, rectified version of the input image, which can then be passed to an OCR engine for more reliable text extraction. On the DocUNet benchmark dataset, UVDoc achieves a Character Error Rate (CER) of 0.179, indicating strong correction performance. The model is also available as an optional preprocessing module within the PP-StructureV3 pipeline, which combines layout detection, OCR, table recognition, seal recognition, and formula recognition for comprehensive document understanding. For developers evaluating this model through gigarouter's API, UVDoc provides a single-call solution for document rectification, with no need to manage dependencies or hardware. It is suitable for any application where document images are captured under uncontrolled conditions and require normalization before text extraction.